This collection of Machine Learning Statistics brings together the latest data on market growth, regional adoption, funding, enterprise use cases, industry applications, technologies, and the ML job market. Drawing primarily on research and industry data from 2024 to 2026, these figures provide a clearer picture of where machine learning stands today and how the field is evolving.

Machine Learning Market Size & Growth

• According to a market analysis from The Business Research Company, the global machine learning market stood at $93.73 billion in 2025 and is on track to hit $127.94 billion in 2026, growing at a CAGR of 36.6% from 2026 to 2030. 
• The same report projects the market climbing steadily to roughly $175 billion by 2027, $240 billion by 2028, and nearly $320 billion by 2029, before reaching an estimated $445.25 billion by 2030, more than 4.7 times its 2025 valuation.

• Research Nester forecasts an even steeper long-term trajectory, estimating the market will expand from $91.31 billion in 2025 to as much as $1.88 trillion by 2035.
• A Market.us study projects the global machine learning market will grow from US$21.5 billion in 2022 to US$582.4 billion by 2032, a CAGR of roughly 39.1% over that decade.
• Grand View Research offers a more conservative estimate, valuing the market at nearly $100 billion in 2025 and projecting growth to $684.4 billion by 2033, at a 26.0% CAGR.
• According to a Wall Street Journal report, advances in AI and machine learning could lift global GDP by 14% by 2030, with gains coming from stronger productivity, more innovation, and leaner operations across most sectors of the economy.
• Accenture and Frontier Economics’ 2017 analysis projected that AI could raise corporate profitability by an average of 38% by 2035, adding roughly $14 trillion in economic value worldwide. 

Machine Learning Market by Region

• Statista data shows the United States as the largest national machine learning market, valued at $21.14 billion in 2024, while the global market is projected to grow at a 36.08% CAGR between 2024 and 2030, reaching a market volume of $503.4 billion in 2030. 
• AIPRM data indicates that in 2024, the United States held the highest projected machine learning market size globally, followed by China, Germany, and the UK.
• Grand View Research reports that North America was the largest regional market in 2025 with a 30.3% revenue share, while Asia-Pacific is projected to record the highest CAGR from 2026 to 2033, making it the fastest-growing region.
• According to Business Research Insights, North America led the global AI/ML market with a 42% share in 2026, a position attributed to strong technology infrastructure and supportive government AI initiatives.
• The Statista dataset expects Asia’s machine learning market to grow from $18.65 billion in 2023 to $184.6 billion by 2030.

• Statista forecasts Europe’s machine learning market will rise from $43.4 billion in 2023 to $144.62 billion by 2030.

Adjacent AI/ML Market Segments

• Market.us projects the global Machine Learning Operations (MLOps) market will grow from $2.98 billion in 2024 to $75.42 billion by 2033, expanding at a CAGR of 43.2% over the period.
• The machine learning-as-a-service (MLaaS) segment is forecast to rise from $45.76 billion in 2025 to $209.63 billion by 2030, reflecting a 35.58% CAGR, according to Mordor Intelligence.
• Statista data shows the broader global AI market valued at $260 billion in 2025, with projections pointing to $1,200 billion by 2030.
• The global explainable AI market is projected to reach $24.58 billion by 2030, per estimates from NMSC.
• Precedence Research forecasts the global natural language processing market expanding from $42.47 billion in 2025 to $791.16 billion by 2034.
• The autonomous AI agent market is expected to grow from $8.5 billion in 2026 to $35 billion by 2030, according to Deloitte.
• Grand View Research projects the global computer vision market will surpass $58 billion by 2030.
• Market.us projects the global AI hardware market will reach $87.68 billion by 2026.
• SQ Magazine reports that the global cloud computing market stands at $1,106.28 billion, with the cloud ML segment projected to grow at a 28.3% CAGR through 2033, while MLOps cloud deployments hold a 51% market share.
• Separately, the cloud ML platform market is expected to grow at a 6.9% CAGR between 2026 and 2033.
• Business Research Insights values a narrower slice of the market, the machine learning and AI space focused on automotive, agriculture, and manufacturing applications, at $17.02 billion in 2026, projecting growth to $238.24 billion by 2035 at a 33.6% CAGR.

Funding & Investment in Machine Learning

Top-Funded Companies & Platforms

• According to Statista, OpenAI ranked as the top-funded machine learning platform as of June 2024, with accumulated investment reaching $11,300.01 million.
• The same dataset places Scale AI and Adept as distant runners-up, with $602.6 million and $415 million in funding, respectively.
• Bloomberg reported that OpenAI closed a $6.6 billion funding round in October 2024, pushing its valuation to $157 billion and cementing its position as the most heavily funded machine learning company globally.

Funding by Sector & Region

• SQ Magazine’s 2026 report puts global funding for ML startups at $98.7 billion, of which U.S.-based startups secured $45.2 billion, and places the average Series B valuation for ML firms at $245 million. 
• SQ Magazine’s 2026 funding data indicates that ML-focused cybersecurity startups brought in $10.9 billion, a 55% jump in funding.
• The same report shows healthcare-oriented ML ventures drawing $13.4 billion, with the money aimed at diagnostics.
• VC-backed ML startups reached a median exit valuation of $620 million, according to the report.
• Corporate arms of Big Tech companies supply roughly 22% of all ML investment.
• Climate-focused ML startups secured $3.1 billion to develop sustainability technology.
• Europe accounts for 18% of global ML funding, with fintech leading the region’s investment.

Enterprise Adoption & Challenges

Adoption Levels & Enterprise Use

• IBM’s Global AI Adoption Index, released in January 2024, found that 42% of enterprise-scale organizations (over 1,000 employees) had actively deployed AI, while another 40% remained in the exploration and experimentation phase.
• According to the 2025 IBM CEO Study, 61% of CEOs report their organizations are actively adopting AI agents and preparing for large-scale deployment.
• U.S. Census Bureau data shows AI adoption among businesses rising from 3.7% in September 2023 to 5.4% by February 2024, with the information sector leading adoption at 18.1%.
• SQ Magazine’s 2026 industry outlook reports that enterprise access to AI tools expanded 50% year-over-year, growing from under 40% to under 60% of the workforce. It also projects that large enterprises will command 55.61% of the machine learning market share in 2026.
• By enterprise size, Grand View Research’s segmentation shows large enterprises as the largest group within the machine learning market, holding a 63.8% share in 2025.
• IDC forecasts that by 2025, Global 2000 organizations would allocate 40% of their core IT spending to AI initiatives.
• SQ Magazine also reports that 81% of Fortune 500 companies now apply machine learning to core enterprise functions such as customer service, supply chain management, and cybersecurity.
• Within legal and compliance teams, ML-powered document automation has been deployed by 44% of organizations, per the same 2026 report.
• Machine learning is now embedded in 72% of ERP systems, primarily to automate invoice processing and vendor performance tracking.
• Business Research Insights reports that 46% of enterprises invested in AI solutions in 2026 to improve operational efficiency and predictive analytics capabilities.

Enterprise Use Cases

According to SQ Magazine’s 2026 use-case data on enterprise ML adoption, companies are applying machine learning across a broad range of business priorities, with cost efficiency topping the list. The most common use cases break down as follows:

• 42% of enterprises put cost reduction first, using ML optimization to trim operating expenses.
• 39% rely on ML to uncover customer insights and drive predictive intelligence.
• 36% use ML models to deliver more personalized, higher-quality customer experiences.
• 33% of organizations use ML workflows to automate internal processes.
• 31% apply predictive analytics to strengthen customer retention.
• 30% run chatbots and support tools on conversational ML.
• 29% adopt ML for recommender systems, and another 29% for fraud prevention.
• 28% turn to ML to curb churn and support customer acquisition.
• 27% forecast demand to improve inventory and logistics planning.
• 22% power loyalty programs with ML-driven engagement.

Adoption Challenges

• The same source identifies the leading obstacles to enterprise ML adoption in 2026 as scaling models into production (46% of organizations), model versioning and reproducibility (44%), and a lack of senior management buy-in (37%).
• Additional barriers include cross-framework integration challenges (36%), duplicated efforts across distributed teams (31%), data quality and labeling inconsistencies (29%), MLOps talent shortages (27%), and regulatory compliance concerns (25%).
• Data privacy concerns held back full deployment for 28% of organizations, particularly in sensitive industries, according to Business Research Insights.
• Zippia reports that 89% of business leaders expect AI to improve employee performance and productivity by 2035, yet only 55% of employees share that outlook.

Machine Learning Across Industries

• SQ Magazine’s 2026 industry breakdown shows BFSI leading ML adoption at 20% usage for fraud detection and risk assessment, followed by automotive (16%, driven by autonomous driving systems), healthcare (14%, diagnostics and patient care), retail (13%, personalization and supply chain), manufacturing (12%, predictive maintenance), advertising and media (11%, targeting campaigns), and energy (9%, grid optimization and sustainability).

Banking & Finance

• The global AI market within BFSI is projected to grow at a 29.6% CAGR between 2022 and 2028, reaching $15.32 billion by the end of the forecast period.
• Analysts projected AI platform revenues within the insurance sector would grow 23% between 2019 and 2024, reaching $3.4 billion.
• SQ Magazine’s 2026 data shows that ML-enhanced predictive models now perform 38% of finance-department forecasting tasks.
• Allied Market Research projects the machine learning market within banking and finance will reach $21.27 billion by 2031.
• Autonomous NEXT, in a report dating back to roughly 2018, projected that AI could cut operating costs across financial institutions by 22% by 2030, translating into potential savings of up to $1 trillion.
• The Commodity Futures Trading Commission (CFTC) reported that in 2024, 99% of leading financial services firms in derivatives markets had deployed AI in some capacity, primarily for risk management, fraud detection, and compliance.

Healthcare

• In 2026, 75% of U.S. health systems are reported to be using or planning to adopt at least one AI application, per SQ Magazine.
• More than 80% of health system and health plan executives expect generative and agentic AI to deliver moderate-to-significant value in 2026.
• Precedence Research forecasts the global AI and ML market in healthcare reaching approximately $613.81 billion by 2034.
• The global healthcare chatbot market is projected to reach roughly $498.5 million by 2026, according to Zipdo.

Retail, Sales & Marketing

• SQ Magazine’s 2026 enterprise data shows that 55% of enterprise CRM platforms now integrate ML-based sentiment and churn-analysis tools, while ML-driven inventory optimization has reduced stockouts by an average of 23% for large retailers.
• The same report notes that 29% of B2B SaaS companies now offer ML-based personalization as a core service feature.
• Retail businesses using machine learning saw sales and profit growth of 14.2% and 8.1%, respectively, between 2023 and 2024, compared with notably slower growth of 6.9% and 3.1% at companies that did not adopt ML tools over the same period.
• Grand View Research found that the advertising and media segment held the largest end-use market share, at 17.3%, in 2025.

Other Industries

• According to SQ Magazine’s 2026 enterprise applications data, HR departments at large U.S. enterprises now use machine learning in 61% of recruitment and talent-scoring workflows.

• SQ Magazine’s 2026 enterprise applications report found that ML-powered cybersecurity tools blocked 34% more threats than traditional systems over the preceding year.

• SQ Magazine’s 2026 enterprise data shows that ML-powered chatbots now handle more than 60% of tier-1 customer queries without human escalation.

Machine Learning Technologies & Platforms

Framework & Library Market Share

• Framework-share data for 2026 shows TensorFlow holding a commanding 41.74% of the machine learning framework market, followed by PyTorch at 25.90%, OpenCV at 17.88%, and Keras at 14.48%.
• Together, TensorFlow and PyTorch account for more than 67% of total framework usage, with TensorFlow’s lead over PyTorch standing at nearly 16 percentage points, underscoring how concentrated the top of the market has become, even as tools like OpenCV and Keras continue to serve specialized niches.

Emerging Technology Trends

• SQ Magazine’s 2026 outlook highlights several emerging technology trends: TinyML adoption grew 33%, fueled by smart-home and industrial IoT use cases, while quantum machine learning has reached proof-of-concept stage at 8% of large research labs.
• The same report notes that synthetic data libraries expanded to include 12 new government-backed repositories, reinforcement learning enterprise trials rose 28% (particularly in robotics and logistics), and energy-efficient ML training frameworks have been adopted by 37% of organizations operating under ESG mandates.
• Cross-lingual machine learning models now translate with more than 91% accuracy across over 80 languages, according to the same 2026 data.

Infrastructure & Cloud Platforms

• Data centers could consume more than 1,000 terawatt-hours of electricity in 2026, driven in part by AI and ML workloads, SQ Magazine reports.
• On the infrastructure side, 95% of new digital workloads are now built on cloud-native platforms, AWS maintains a 31–32% share of cloud infrastructure leadership, and hybrid or multi-cloud setups have been adopted by 98% of businesses.
• Grand View Research found that the services segment held the largest component-based market share, at 55.2%, in 2025.

Market Segmentation & Competitive Landscape

• Business Research Insights reports that the integration of natural language processing and computer vision has driven 33% growth in advanced AI application usage, with the top five industry players accounting for 55% of the market through partnerships, acquisitions, and platform innovation.
• The same source notes that deep learning technology captured 39% of ML deployments in 2026, concentrated in image recognition, autonomous systems, and predictive analytics.
• It further highlights that in 2024, AI platforms incorporating explainable AI and automated model training saw adoption increase by 27% globally.

Machine Learning Jobs, Skills & Salary

Job Growth & Demand

• AI-related job postings reached 4.2% of all Indeed listings by late 2025, with that momentum carrying into 2026, according to SQ Magazine.
• The same report finds that nearly 45% of U.S. data and analytics job postings now reference AI-related skills.
• SQ Magazine’s Q1 2026 data shows the U.S. ML job market growing 32%, led largely by demand for applied ML engineers, while MLOps positions have surged 45% amid rising production-deployment needs.
• University ML program enrollments rose 23% across top U.S. institutions, and freelance ML contracts increased 40% on specialized job platforms, per the same 2026 dataset.
• Tech, finance, and healthcare sectors collectively posted 48,000 ML jobs year-to-date in 2026.
• Global employment for machine learning engineers is projected to grow 22% between 2020 and 2030.
• The U.S. Bureau of Labor Statistics projects computer and information science jobs expanding from 33,500 in 2021 to 40,600 by 2031 — a 21% growth rate, compared with just 5% projected growth across all occupations over the same period, plus an estimated 3,300 additional openings per year to replace workers leaving the field.

Salary & Skills Gap

• SQ Magazine’s 2026 job-market data puts the median ML engineer salary at $168,000, with top roles exceeding $220,000.
• The same report finds that Python, PyTorch, and TensorFlow dominate 87% of ML job requirements, and 62% of ML roles require experience with cloud platforms such as SageMaker or Vertex AI.
• Women now make up 29% of the ML workforce, up three percentage points year-over-year, while bootcamp graduates fill 20% of new ML roles at scaling tech firms, per SQ Magazine’s 2026 findings.

Methodology 

This roundup compiles machine learning statistics from a wide range of research articles, industry reports, and statistics publications, most of them published between 2024 and 2026, with a small number of earlier studies included where their forecasts extend to 2024 or later or where they provide useful context. The sources include market forecasts from research firms such as The Business Research Company, Grand View Research, and Business Research Insights, enterprise and executive surveys from organizations like IBM and IDC, government and regulatory data from the U.S. Census Bureau and the Commodity Futures Trading Commission, and funding and industry data reported by Statista, Bloomberg, and SQ Magazine.

Every data point was reviewed for accuracy and checked for duplication, and entries that repeated or contradicted one another within a section were removed. The remaining figures were paraphrased and attributed to their source, along with the year they refer to. Research firms often size the same market differently. Because of this, diverging estimates, such as the global market value in 2025, are shown with their individual sources rather than merged into one number. When a source does not state the period a figure covers, the data is attributed to the publication and its edition year instead. This article reflects Memeburn’s independent analysis of publicly available data.

References

• A3Logics (2025). Machine Learning Statistics That Matter in 2025 – Market Insights & Future Trends.  A3Logics. Available at: https://www.a3logics.com/blog/machine-learning-statistics/.
• Itransition.com (2026). Itransition. Available at: https://www.itransition.com/machine-learning/statistics.
• McCain, A. (2023). 25+ Incredible Machine Learning Statistics [2026]: Key Facts About The Future Of Technology.  Zippia. Available at: https://www.zippia.com/advice/machine-learning-statistics/.
• Pangarkar, T. (2026). Machine Learning Statistics By Programs, Technology, Performance (2026).  Market.us Scoop. Available at: https://scoop.market.us/top-machine-learning-statistics/.
• Ramirez, S. (2025). Machine Learning Statistics 2026: Growth Secrets.  SQ Magazine. Available at: https://sqmagazine.co.uk/machine-learning-statistics/.
• Revankar, S. (2024). Machine Learning Statistics By Market Size, Region, Models And Usage.  Electro IQ. Available at: https://electroiq.com/stats/machine-learning-statistics/.
• Sanghavi, A. (2024). 50+ Machine Learning Statistics That Matter in 2024.  G2dotcom. Available at: https://learn.g2.com/machine-learning-statistics.
• Wasay, A. (2026). Machine Learning Market Outlook and Forecast Report 2026-2030.  The Business Research Company. Available at: https://www.thebusinessresearchcompany.com/report/machine-learning-global-market-report.
• www.grandviewresearch.com (n.d.). Machine Learning Market Size, Share | Global Industry Report, 2025. Available at: https://www.grandviewresearch.com/industry-analysis/machine-learning-market.

The post Machine Learning Statistics 2026: The Complete Data Roundup appeared first on Memeburn.

What PewDiePie Is Releasing With Ajax

Ajax is a 9-billion-parameter model fine-tuned from Alibaba’s Qwen 3.5 9B and designed to handle everyday agent tasks inside Odysseus, such as browsing the web, running private searches, and managing email and calendars without sending data to a cloud provider. According to PopularAI’s technical breakdown, he first trained it on successful Odysseus task traces, then spent weeks running GRPO reinforcement learning on tool-use workflows, and the full-precision version needs roughly 22GB of video memory, with an FP8 build at around 11GB and smaller quantised versions planned. The model is scheduled to go live on October 3 at 08:25 Japan time, which is 7:25 pm Eastern on October 2, and no license or independent benchmark had been published at the time of writing.

Ajax extends a project that began with Odysseus, the free workspace PewDiePie released at the end of May, which Memeburn covered when it launched as a self-hosted alternative to handing personal data to big tech.

Why OpenAI Banned Him, Twice

In the video, PewDiePie said that OpenAI “banned me twice” while he was building the model, and TechCity Authority reported that he linked the bans to his attempts to distill reasoning tokens from OpenAI’s models. Posts on X identified the model as Sol, the flagship OpenAI released in July as part of its GPT-5.6 family, as Decrypt detailed at launch. According to Free Press Journal, the first ban was reversed on appeal before a second one followed, and OpenAI’s consumer terms prohibit extracting output programmatically. OpenAI has not commented publicly on either ban.

The reasoning tokens matter because OpenAI does not show its models’ full chain of thought to users, and that hidden reasoning is exactly what makes a large model valuable as a teacher for a smaller one. One widely shared post claimed he had used an exploit to decrypt those tokens, and such a technique does exist, since researchers disclosed in August that encrypted reasoning blocks from OpenAI, Anthropic and Google could be replayed into a weaker model and read back in plain text, a flaw the providers have since patched on their servers. No source has shown that PewDiePie used that method, however, and he has not said so himself.

A Pipeline the US Government Has Already Described

On September 8, the NSA, CISA and FBI published a joint advisory accusing DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI of distilling US frontier models, and as Unite.AI summarised, the signals it told providers to watch for included jailbreak prompts aimed at extracting chain-of-thought reasoning. Taking a leading American model’s reasoning and using it to improve a fine-tune of a Chinese open-weight model is, in outline, the pipeline that advisory describes, and it is also roughly what PewDiePie says he was attempting from a home GPU rig.

The scale is obviously not comparable, and nothing suggests that a single creator building a 9B model for his own workspace poses the kind of risk the advisory was written about. The point is that a provider’s detection system does not see scale or intent at the moment it flags an account, because it sees request patterns, and the pattern of someone pulling reasoning out of a frontier model to train another model looks much the same whether that someone runs a lab in Beijing or a YouTube channel in Japan. As Memeburn noted when the advisory was published, the same signals can also catch ordinary developers whose usage happens to resemble a distiller’s.

The Ban Was the Visible Option

The most consequential part of the advisory was not the list of names but its recommendation that providers respond to suspected distillation by quietly degrading or altering the answers they serve, instead of blocking accounts outright, and that they avoid telling suspected users this is happening. PewDiePie received the opposite treatment, since a ban is the one response a user can see, appeal and talk about in a video watched by millions.

Had OpenAI applied the approach the advisory recommends, he might simply have received weaker outputs, trained Ajax on them and never known why the results were worse. A famous creator can turn a ban into a story, but an ordinary developer whose answers are silently degraded has no ban to point to, which makes his case a useful reminder that the quieter policy is also the one that leaves no trace for anyone to contest.

The Hypocrisy Argument Only Goes So Far

Much of the reaction on X took his side on principle, with one popular post arguing that companies which trained their models on the public’s data are in a weak position to object when a user takes something back. That argument resonates because the legality of training on scraped data is still being fought over in courts, but it mixes two different things, since PewDiePie agreed to OpenAI’s terms when he opened an account, while the people whose writing ended up in training sets never signed anything. Enforcing a contract a user accepted is a different matter from the unresolved question of what AI companies owed the authors of their training data, even if both disputes are about who gets to learn from whose work.

Removing Refusals Is the Part Worth Watching

The less discussed part of the video is that Ajax has had its refusal behaviour stripped out using Heretic, an open-source tool that applies directional ablation to weaken the internal patterns associated with declining requests. PewDiePie compared the process to brain surgery and said he drew his own line at harm to other people or to the user, keeping some safeguards in place, although PopularAI notes that ablation can also reduce output quality. Once Ajax’s weights are public, his particular line becomes a default rather than a limit, because anyone downloading an open-weight model can run the same kind of tool again and remove whatever restrictions remain.

What to Check When Ajax Goes Live

The launch will settle several questions that the video left open, and these are the details worth looking for on day one:

• Whether Ajax ships under a stated license, since the Qwen base model uses Apache 2.0 but no terms for Ajax itself have been announced.
• Whether PewDiePie discloses how much of the final training data came from OpenAI outputs, given that he has described them only as part of a seed-data process alongside Odysseus traces and synthetic data.
• Whether independent testers can reproduce the roughly nine-in-ten task completion rate he reported internally.
• Whether OpenAI says anything about why the account was banned, which would show whether the bans were about reasoning extraction specifically or about automated use in general.

FAQs

Why did OpenAI ban PewDiePie?

PewDiePie said OpenAI banned him twice while he was building Ajax, and he linked the bans to his attempts to distill reasoning tokens from OpenAI’s models, which the company’s terms prohibit, although OpenAI has not commented on the case.

What is PewDiePie’s Ajax AI model?

Ajax is a 9-billion-parameter local AI model fine-tuned from Alibaba’s Qwen 3.5 9B to run agent tasks such as browsing, search, email and calendar management inside PewDiePie’s self-hosted Odysseus workspace.

When is Ajax released?

Ajax is scheduled to go live on October 3, 2026 at 08:25 Japan time, which is 7:25 pm Eastern Time on October 2.

What is AI model distillation?

AI model distillation is the practice of training a smaller model on the outputs of a larger one so that the smaller model inherits some of its abilities, and frontier labs prohibit it when the outputs come from their models without permission.

Did PewDiePie use an exploit to decrypt OpenAI’s reasoning tokens?

There is no evidence that he did, because the claim comes from posts on X rather than from PewDiePie, even though researchers disclosed in August that encrypted reasoning blocks could be extracted through a technique that providers have since patched.

The post PewDiePie’s OpenAI Ban Shows Distillation Rules Can’t Tell a YouTuber From an AI Lab appeared first on Memeburn.

First, What Is “Custody” and Why Does It Matter?

Custody just means who actually holds your assets. With stocks, a big bank or brokerage keeps them for you. With crypto, it’s trickier because owning crypto really means controlling a private key, a secret code that works like the only password to a vault. Whoever holds the key can move the money, and there’s usually no undo button.

Investment advisers are the professionals who manage money for clients, often for a fee. For years, U.S. rules said they had to keep client assets with a “qualified custodian,” typically a bank or regulated broker. The problem? For many crypto tokens, no such custodian existed. SEC Chairman Paul Atkins put it bluntly: “For too many assets a qualified third-party custodian simply does not exist yet.”

So advisers mostly stayed away, or pointed clients to Bitcoin ETFs instead. That’s a big reason so much traditional money has flowed into crypto through funds, like the inflows behind Bitcoin ETFs’ best week of 2026, rather than through direct holdings.

What the SEC Crypto Custody Proposal Actually Changes

The SEC’s official announcement covers registered investment advisers, registered funds like mutual funds and ETFs, and business development companies (BDCs), which are investment firms that lend to or invest in smaller businesses. It opens two new paths.

It also tidies up older rules. The main adviser custody rule would move from Rule 206(4)-2 to a new Rule 223-1, and firms could keep certain records directly on crypto networks, as long as they can produce them in human-readable form.

Path 2 is designed as a bridge, not a destination. The quarterly check means an adviser holding a niche token has to keep looking for a proper custodian. That’s sensible. It lets advisers serve clients now without pretending they’re as well-equipped as specialist custodians.

The “Self-Custody” Label Is a Bit Misleading

Here’s a detail that’s easy to miss. In crypto, “self-custody” usually means you hold your own keys, often in a hardware wallet. This proposal isn’t about that.

Commissioner Hester Peirce made the point herself, putting “self-custody” in quotation marks in her statement on the proposal. What the rule really describes is adviser custody. She has long argued that regulators should zealously protect people’s right to hold their own crypto, and this rule doesn’t change that right either way.

Why does the difference matter for you? Because if your adviser holds your crypto under Path 2, the adviser becomes the vault. Your money’s safety depends on how well that firm manages keys.

2026’s Hacks Show Why the Safeguards Matter

The proposal’s conditions read like a list of lessons from this year’s crypto disasters. Exchange Bitget lost around $350 million in a hack and then leaned on a protection fund that’s really a Bitcoin bet. Earlier, the Coldcard wallet hack drained $114 million in Bitcoin and shook confidence in hardware wallets.

Requirements like two-person approval and separate addresses for each client are designed to stop exactly those failures. Two-person approval means one rogue employee or one stolen laptop can’t move funds alone. Separate addresses mean one client’s assets can’t be quietly mixed with another’s.

Rules can require good practices, but they can’t make a small advisory firm as hard to hack as a specialist custodian. Smaller advisers may find the cost of independent accountant reports and cybersecurity reviews high enough that they skip crypto entirely, or stick with ETFs. That might be the most realistic outcome for many.

The Real Weak Spot: A Two-Person SEC With No Law Behind It

This is where we think most coverage stops too early. The SEC normally has five commissioners, with no more than three from one party. Peirce leaves today, October 2. That leaves Chairman Atkins and Commissioner Mark Uyeda, both Republicans, with three seats empty. SEC rules let whoever’s in office act as a quorum, so the agency can keep going.

That keeps the proposal moving, but it also means the rule could be finalized with no dissent on the record and no Democratic input. Agency rules are easier to change than laws. A future SEC with a different majority could reopen and rewrite this one through its own rulemaking. That’s not hypothetical: the SEC’s 2023 custody proposal, which would have made crypto custody much harder, was later withdrawn after heavy industry pushback.

A law from Congress would be far sturdier. But the Clarity Act fell short in the Senate on September 15, getting 49 votes when it needed 60. So the SEC is filling the gap one rule at a time. Atkins has framed it as making the U.S. the “Crypto Capital of the World”, but rules built this way can be unbuilt the same way.

With our new proposal, the SEC is taking the most historic step yet to modernize federal securities regulations for crypto assets.

As the Crypto Capital of the World, the U.S. must and will lead. Regulation Crypto Assets will ensure that we do. 🇺🇸 pic.twitter.com/z0MmDF4doV

— Paul Atkins (@SECPaulSAtkins) August 18, 2026

FAQs

What did the SEC change about crypto custody before this proposal? 

The SEC first sent a revised custody plan to the White House for review in late summer. 

Why did the Clarity Act fail in the Senate? 

It fell short of 60 votes after disputes over ethics provisions and stablecoin safeguards. Tax treatment was another sticking point, as seen in Kevin O’Leary’s crypto tax bet on the Clarity Act.

Are banks also moving onto blockchains? 

Yes, from a different direction. In one recent example, 39 U.S. states agreed to build a shared blockchain for banks, aimed at modernizing how banks settle payments.

Are big Wall Street firms already holding crypto? 

Some are, mostly through ETFs. Goldman Sachs, for example, has built large positions, and our report on how Goldman Sachs became the biggest XRP whale on Wall Street shows how institutions are getting exposure.

Can a custodian see everything I do with my crypto? 

Often, yes. Custodians and exchanges keep detailed records for compliance. Privacy tools are evolving, including efforts to bring Zcash-style privacy to Bitcoin without a fork, though regulators are watching closely.

The post The SEC Wants Your Financial Adviser to Be Able to Hold Your Crypto Keys appeared first on Memeburn.

Where Arbitrum’s $5.45 Million Came From

Arbitrum revenue in September 2026 totalled $5.45 million, of which $4.75 million came from Orbit licensing fees, the payments that chains built with Arbitrum’s Orbit stack make under the Arbitrum Expansion Program. The full breakdown from the dashboard Arbitrum shared shows how far the other lines trail behind.

Arbitrum’s post also said the network had processed 1.5 billion new transactions across the “Arbitrum Platform” so far in 2026, a figure that includes Orbit chains as well as Arbitrum One, and it did not say which earlier month the fivefold increase was measured against.

How Robinhood Chain Pays Arbitrum

Robinhood Chain went live on July 1 as an Orbit chain, and under the Expansion Program any chain that settles outside Arbitrum One and Nova sends 10% of its net protocol revenue back to the ecosystem, with eight percentage points going to the Arbitrum DAO treasury and two to a developer fund. When Standard Chartered initiated coverage of ARB in mid-September, CoinDesk reported that these licensing payments had been worth about $360,000 in July, already 35% of DAO income that month, and that the bank expected roughly $5 million in September, an estimate the final $4.75 million came close to matching.

Most of September’s Money Likely Arrived in the First Two Weeks

The timing of Robinhood Chain’s fees matters because September was not a steady month. On September 1, the chain recorded $3.75 million in daily fees, its fourth consecutive record day, which translated into roughly $370,000 for the Arbitrum ecosystem from a single day of trading. That surge followed the late-August wave of memecoins tied to tokenized stocks that Memeburn covered when a single token cornered most of the supply of tokenized HIMS. By September 11, daily chain revenue had fallen 83% from its peak to about $1.06 million as memecoin traders moved on and gas prices dropped.

The arithmetic suggests the record was front-loaded. Reaching $4.75 million over 30 days required an average of about $158,000 a day for Arbitrum, while the first days of September delivered more than twice that and the level implied by mid-month fees was well below it. Trackers measure Robinhood Chain’s revenue in different ways, some of them including application-layer fees, so this should be read as an estimate rather than a reconstruction, and Arbitrum has not published a daily breakdown that would settle it. The cooling continued into late September, when CryptoTicker reported that protocols on the chain were collecting a weekly average of $5.93 million a day in fees against a 30-day average of $12.44 million.

Robinhood Paid Some of the Gas, and That Ended on September 29

For the chain’s first 90 days, Robinhood covered network fees for swaps made through its own wallet, initially above $5 and, from August 8, above $0.50. As Crypto.news noted, most wallet users in practice paid nothing, and Robinhood never disclosed what the programme cost.

That raises a question the record does not answer, which is how much of the activity behind Arbitrum’s September licensing fees was paid for by Robinhood rather than by traders. The subsidy only applied to Robinhood Wallet users, so bots and other wallets were paying real gas throughout, but no one has published how the fee base split between the two. Since September 30, every transaction on Robinhood Chain pays its own gas in ETH, which makes October the first clean reading of demand that is not partly underwritten by the chain’s owner.

The Tokenization Thesis Is Being Funded by Memecoins

The bull case for Arbitrum’s licensing model, set out most prominently in Standard Chartered’s $10 target for ARB by 2030, rests on traditional assets moving onchain at scale and Arbitrum collecting a share as the infrastructure underneath. The money arriving today comes from somewhere else. KuCoin’s mid-September report put real-world asset trading on Robinhood Chain at under $30 million, less than 0.1% of the chain’s volume, while launchpads such as Pons and memecoins such as CASHCAT drove most of the activity.

None of this makes the thesis wrong, since tokenization was always going to take years, but it means September’s record does not yet count as evidence for it. The regulatory path is also uncertain, because the SEC’s new exemption for tokenized stock trading in the US requires permissioned AMMs and tokens that carry real shareholder rights, conditions that Memeburn found would force several current models to rebuild.

Priority Gas Auctions Are the Number to Watch on Arbitrum One

The most interesting line in Arbitrum’s announcement was not the headline but the claim that Priority Gas Auctions had earned more than $120,000 in their first week. PGA replaced Timeboost on September 24, letting users bid a priority fee on each transaction rather than buying time-based express-lane access, which explains why Timeboost earned only $9,480 in September before it was switched off.

PGA revenue does not appear as a separate line in the dashboard, so it is most likely counted within transaction fees, but its value is that it comes from Arbitrum One itself rather than from a customer chain. If it holds anywhere near its first-week pace, it would become the clearest sign of whether demand on Arbitrum’s own network is growing, independent of whatever happens on Robinhood Chain.

What the Record Means for ARB Holders

The $5.45 million flows to the Arbitrum DAO treasury and developer fund, and ARB holders have no automatic claim on it, since there is no built-in buyback or distribution and any use of the money has to pass through governance. Standard Chartered listed that lack of direct value capture among its own caveats, alongside memecoin dominance and the subsidy expiry, so the record strengthens the DAO’s balance sheet without changing what an ARB token entitles its holder to, unless governance decides otherwise.

FAQs

How much revenue did Arbitrum make in September 2026?

Arbitrum made a record $5.45 million in September 2026, of which $4.75 million, or about 87%, came from Orbit licensing fees paid mainly by Robinhood Chain.

Why did Arbitrum’s revenue jump in September?

Arbitrum’s revenue jumped because Robinhood Chain, which pays 10% of its net protocol revenue to the Arbitrum ecosystem, had its busiest month during a memecoin trading boom, while Arbitrum One’s own transaction fees stayed at about $507,000.

How does Robinhood Chain pay Arbitrum?

Robinhood Chain is built on Arbitrum Orbit and joined the Arbitrum Expansion Program, which requires chains settling outside Arbitrum One and Nova to send 10% of net protocol revenue to the ecosystem, split eight points to the DAO treasury and two to a developer fund.

When did the Robinhood Chain gas subsidy end?

Robinhood’s gas subsidy for Robinhood Wallet users ran for 90 days from July 1 and ended on September 29, 2026, after which every transaction on the chain pays its own gas in ETH.

Do ARB holders receive Arbitrum’s revenue?

ARB holders do not receive Arbitrum’s revenue directly, because the money goes to the DAO treasury and developer fund, and any buyback or distribution would require a governance decision.

The post Arbitrum’s Record Month Was Robinhood’s Memecoin Month. October Is the Real Test appeared first on Memeburn.

What Happened

First, a definition. An AI hallucination is when a chatbot confidently states something false, like inventing a fact, a source, or a conclusion that sounds right but isn’t. It happens because chatbots predict likely-sounding text rather than checking facts against reality.

Here’s the sequence, based on what’s been reported so far:

1. An analyst at U.S. Special Operations Command Pacific, based in Hawaii, fed a chatbot a mix of public information and classified signals intelligence (intercepted communications) about a Chinese ship in the Middle East.
2. The chatbot concluded the ship carried components of a nuclear weapons program headed to Iran.
3. The analyst then used AI again to turn that conclusion into a standard military intelligence report.
4. The report circulated, and forces began preparing to intercept and board the vessel. Aircraft were already in the air.
5. Officials took a closer look, realized the report was AI-generated and wrong, and called off the operation.

A source described the report as “entirely false” and said it “almost started a war.” It’s still not public which chatbot was used, whether it was commercial or government-built, or what the ship was really carrying. The Pentagon and the command involved haven’t publicly explained the incident.

The Real Failure Was Step 3

Most coverage has focused on the hallucination itself. We think the more important moment came afterward.

Chatbots make mistakes all the time. That’s known, and analysts are trained to treat raw tips skeptically. But when the analyst used AI to convert the chatbot’s answer into a standard intelligence report, the guess was put on a uniform. It now looked like every other vetted assessment moving through the system, with the same structure, the same tone, and the same authority.

Think of it as confidence laundering. The chatbot’s uncertainty didn’t survive the formatting step. Nobody reading the finished report could tell that its core claim came from a machine guessing, rather than a human analyst weighing evidence.

There’s a second layer, too. The chatbot mixed open-source information with classified intelligence. Classified material carries extra weight inside the military, because people assume it’s been collected and checked carefully. When an AI blends a shaky public detail with a solid secret one, the shaky part can borrow credibility it never earned.

A former senior U.S. official reportedly described the military’s internal AI tools as “mostly just copies of the commercial stuff wearing lipstick.” If that’s accurate, the same hallucination risks you see in consumer chatbots are now sitting inside intelligence workflows.

People Expect AI to Check Things It Doesn’t

This wasn’t the first time this year that misplaced trust in an AI system ended up in a deadly or near-deadly chain.

On February 28, the first day of the U.S.-Iran war, a U.S. Tomahawk missile hit the Shajareh Tayyebeh elementary school in Minab, Iran, killing well over 100 people, many of them children. The building had once been a military site but had been a school for years. Pentagon investigators have since said that overreliance on Palantir’s Maven targeting system contributed, alongside outdated intelligence data and deep cuts to civilian-harm review staff. Reporting on the Minab strike and the Pentagon’s AI targeting push points to stale human-provided data as the core problem, not an AI malfunction.

🚨🇺🇸🇮🇷 BREAKING:

Multiple media outlets are confirming what Iran’s been saying: A U.S. Tomahawk missile hit the Shajareh Tayyebeh girls’ elementary school in Minab.

Death toll: 165-180, mostly children aged 7-12.

Israel doesn’t possess Tomahawk missiles. Only the… https://t.co/6TXhmROkUn pic.twitter.com/EawEZhOnAL

— Mario Nawfal (@MarioNawfal) March 8, 2026

That’s exactly our point. Users reportedly expected the system to flag outdated records. It wasn’t designed to. Palantir has said it isn’t responsible for the underlying data. In both cases, humans assumed the software was doing a safety check that nobody had actually built in.

“Human in the loop” sounds reassuring, but it only works if the human knows which parts came from the machine and what the machine never checked.

Why the Scale Makes This Urgent

The Pentagon has gone all-in on AI this year. Its GenAI.mil platform launched in December 2025 with Google’s Gemini and has onboarded about 1.7 million unique users, with access extended to more than 3 million people. OpenAI’s ChatGPT Mil and xAI’s Grok for Government joined on September 1. Defense Secretary Pete Hegseth launched an AI acceleration strategy in January.

GenAI.mil runs on unclassified networks, and it isn’t known whether it was involved in this incident. But the scale changes the math. If even a tiny share of 1.7 million users paste AI output into official documents without labels, near-misses like this stop being freak events.

That’s why three Democratic senators, Mark Warner, Jack Reed, and Chris Coons, asked for an investigation on September 19, including unrestricted inspector general access. They acknowledged AI has legitimate intelligence uses but argued that standard safeguards must stop AI errors from prompting military action against nuclear-armed countries. Warner had already warned days earlier that “we cannot wait for AI companies to self-regulate”.

We cannot wait for AI companies to self-regulate. They are raising every alarm. Congress must act this week to meet the rapidly developing threat landscape posed by AI. Let’s take bipartisan action NOW to meet this moment.

— Mark Warner (@MarkWarner) September 14, 2026

Superintelligence vs Everyday Errors: A False Choice

Much of the AI safety conversation focuses on future systems smarter than humans. Some researchers have walked away from top labs over it, like the Anthropic researcher who warned AI could kill everyone by 2030.

This incident suggests a more immediate danger: ordinary AI, used carelessly, inside high-stakes systems. We don’t think it’s either-or. But the near-miss shows that a fairly mundane failure, a chatbot inventing a conclusion plus an AI tool dressing it up, can push two nuclear powers toward confrontation without any superintelligence required.

It also raises a question that matters for international stability. If a U.S. boarding had gone ahead, China would have had no way to know the trigger was an AI hallucination. As militaries in the U.S., China, and elsewhere race to use AI, from analysis tools to humanoid robots in military training, mistakes can travel faster than any hotline.

Three Lessons for Anyone Using AI at Work

You probably won’t ever brief a general, but the same failure happens in offices every day. A chatbot summary gets pasted into a slide deck, then the deck gets forwarded, and three emails later nobody knows the key number came from AI.

• Label what AI wrote. A simple note like “AI-drafted, not verified” keeps the uncertainty attached to the claim.
• Check the claim, not the polish. Clean formatting and confident language are exactly what chatbots are best at. They’re not evidence.
• Trace the source. If an AI conclusion rests on a mix of sources, find out which one is carrying the weight before you act on it.

Schools are already struggling with this, as AI in education statistics show many universities can’t reliably detect AI-written work. If teachers can’t spot it, busy colleagues won’t either.

The chatbot’s hallucination was predictable. What turned it into a near-war was a process that let a machine’s guess become an official-looking report without a label, plus a culture that assumes software checks things it doesn’t. Better AI models will help, but clear labeling and a habit of verification would have caught this faster, and they cost almost nothing.

FAQs

Why do AI chatbots hallucinate? 

Chatbots generate text by predicting likely next words, not by checking a database of facts. Newer models hallucinate less, but none are error-free.

Who checks whether AI models are safe? 

Mostly the companies themselves, plus a growing set of outside evaluators. Anthropic recently chose a consulting firm for that job, and our look at how Anthropic picked Accenture as its first AI watchdog explains why that choice raised eyebrows.

How is AI changing the defense industry? 

Startups now build AI-driven surveillance and drones that governments buy at scale. Drone maker valued at $6.4 billion sells “eyes in the sky” rather than weapons.

Did the U.S.-Iran conflict affect financial markets? 

Yes. Strikes and escalation fears hit risk assets, including crypto

Is the U.S. government regulating AI safety? 

There’s no comprehensive federal AI law yet. Policy so far has come from executive actions and industry deals.

The post A Chatbot Almost Started a War With China. The Scariest Part Wasn’t the AI’s Mistake appeared first on Memeburn.

What Siri Recap and Live Rewind Actually Do

Siri Recap is an opt-in Apple Watch feature that listens to conversations throughout the day and turns them into short, high-level summaries with a title and key points, which the owner can read later in the Siri app. Live Rewind is the on-demand companion, keeping a rolling 15-second buffer of nearby speech and showing it as text when the wearer double-presses the Digital Crown, which is useful for catching a name or a number you just missed. According to Apple’s support page for Audio Intelligence, the same family also includes Sound Recognition, which flags sirens, alarms and doorbells, and automatic song identification through Shazam.

Both conversation features arrive as a beta later in 2026, in English only, and they require an Apple Watch Series 12 or Ultra 4 paired with an iPhone 16 or newer, excluding the iPhone 16e. That last requirement undercuts the idea that Apple is simply switching on an AI companion for the many millions of watches already on people’s wrists, since what it has actually built is a reason to upgrade, and the privacy questions will grow with the installed base rather than arriving all at once.

Where the Audio Goes and What Apple Keeps

Apple’s answer to the data question is detailed and, for the most part, convincing. Audio is routed into a protected area of the watch’s S11 chip that the operating system, apps and Apple itself cannot access, and it is deleted once it has been analysed. For Siri Recap, MacRumors reported that an encrypted, condensed text version less than half the length of the original is then sent to Apple’s Private Cloud Compute servers to generate the summary, which means the feature is not purely on-device even though no recording ever leaves the watch.

The table below sets out what each Audio Intelligence feature hears, what it keeps, and what anyone nearby would notice.

Apple also says Siri Recap does not attribute what was said to individual speakers and tries to filter out sensitive material such as financial details and authentication data, although its own documentation concedes that summaries may be inaccurate and that the filtering may miss things.

Apple’s Privacy Design Answers the Wrong Question

Every one of those safeguards is aimed at the same threat, which is someone other than the wearer getting hold of the wearer’s data. Jennifer King, a privacy fellow at Stanford’s Institute for Human-Centered AI, told Yahoo Tech that Apple appears to have worked hard to avoid creating an audio trail that law enforcement could request, and Apple itself states that there is no recording to hand over because none exists. In the same report, Thorin Klosowski of the Electronic Frontier Foundation judged that the wearer’s privacy had been handled reasonably while the privacy of bystanders was “not really considered.”

That is the distinction the whole debate turns on. Encryption, deletion schedules and a sealed-off chip determine who can see a conversation after it has been captured, but none of them determine whether the other people in that conversation agreed to have it captured at all. On that second question, Apple’s only guidance is that wearers should “be mindful of those around you” when conversations might be private or sensitive, which places the entire burden on the person who chose to switch the feature on.

A Chime After the Fact Is Not Consent

Apple’s main concession to bystanders is that Live Rewind plays an audible chime from the watch speaker when it is triggered, even if the watch is in silent mode or connected to headphones. The problem is the order of events, because by the time the chime sounds, the previous 15 seconds have already been captured and turned into text, so the signal tells the other person what has happened rather than giving them a chance to object. As Six Colors pointed out, someone caught by Live Rewind has no way to say no.

Siri Recap, which is the feature that actually listens all day, does not even offer that much, since it gives people nearby no indication at all that their conversation is being summarised.

The Watch Is the Hardest Device to Notice

This is where the form factor matters more than the technology. People have learned to read a raised phone as a camera and to treat smart glasses with some suspicion, and Klosowski drew exactly that contrast, noting that an Apple Watch with these features looks no different from one without them and carries no visible recording light of the kind Meta’s glasses use. A watch is the one gadget most people have stopped seeing as a gadget, which makes it the least likely place for anyone to expect a microphone that is paying attention.

That is also why the risk extends beyond Apple. If always-on summarisation becomes normal on the most widely worn smartwatch, it sets the expectation for every wearable that follows, and the social rule that a conversation is not being captured unless you can see something recording will quietly stop being true.

Is Siri Recap Legal Where Everyone Must Consent to Recording?

Several US states, California among them, require the consent of everyone in a private conversation before it can be recorded, and it is not yet clear how those laws apply to a feature that never produces a recording. King’s view is that Apple’s design avoids being wiretapping in the strict sense by producing neither an audio file nor a verbatim transcript, only a summary, but that is an interpretation rather than a ruling, and no court has yet tested it. Memeburn is not offering legal advice here, and anyone who plans to use Siri Recap in meetings, interviews or other sensitive settings should check the rules where they live and work.

How to Turn Off Siri Recap and Live Rewind

Because every Audio Intelligence feature is opt-in, a new Series 12 or Ultra 4 will not listen to conversations unless its owner switches the features on. For anyone who has enabled them and wants to scale back, Apple’s support documentation describes the controls as follows:

• Siri Recap can be switched off in the Apple Watch app on iPhone, paused from Control Center on the watch, or scheduled so that it only listens at certain times or places.
• Live Rewind is disabled by turning off the Double Click Digital Crown option in the Siri section of the Apple Watch app.
• Sound Recognition is turned off under Accessibility settings, and automatic song identification stops once its Smart Stack widget is removed.

Saved summaries and snippets can also be deleted at any time, and the deletion applies across every device signed in to the same Apple account.

FAQs

Is the Apple Watch always listening?

Only if the owner turns on Siri Recap, which is an opt-in feature on the Apple Watch Series 12 and Ultra 4 and can be scheduled to run at set times, while Live Rewind keeps a rolling 15-second buffer only when it has been enabled.

Does Siri Recap record audio?

No, because Apple says audio is analysed inside a protected area of the S11 chip and then deleted, after which a condensed text version is sent to Private Cloud Compute to create a summary that is kept for seven days unless saved.

Will people nearby know if my Apple Watch is using Siri Recap?

They will not, since Siri Recap gives no indication to anyone nearby, whereas Live Rewind plays an audible chime when activated, although only after the previous 15 seconds have already been captured.

Which Apple Watch models support Siri Recap?

Siri Recap and Live Rewind require an Apple Watch Series 12 or Ultra 4 paired with an iPhone 16 or later, excluding the iPhone 16e, and both launch as an English-only beta later in 2026.

How do I turn off Siri Recap on Apple Watch?

You can turn Siri Recap off in the Apple Watch app on iPhone, pause it from Control Center on the watch, or set a schedule so that it only listens during particular hours or in particular places.

The post Apple Watch’s Siri Recap Protects Your Data, but Nobody Asked the Person Across the Table appeared first on Memeburn.

Rogbid Loop W Specs at a Glance

Some quick definitions first. SpO2 is your blood oxygen level. HRV (heart rate variability) is the tiny variation in time between heartbeats, a common signal for stress and recovery. ECG (electrocardiogram) records your heart’s electrical activity, which can reveal irregular rhythms.

How Rogbid Gets to $60

Rogbid isn’t a household name, but it’s moving fast. The Loop W is part of a family of screenless bands the brand has rolled out since spring. The original Loop launched in late March at $69.99, the Loop V followed in August at $69.99 with fall detection, and the simpler Loop N sits at $49.99. The Loop W now slots in the middle at $59.99.

The parts list explains the price. The JL7073 chip comes from JieLi, a supplier that shows up in plenty of budget wearables. The Goodix heart-rate sensor is a known, affordable component. Rogbid’s real cost advantage, though, is the business model: there’s no subscription, no AI coach, and no cloud service to fund every month.

That’s the whole pitch. Rogbid isn’t trying to out-science Whoop. It’s betting that plenty of people want their sleep and heart data without paying rent on it.

The ECG Question: Feature vs Proof

This is the part we’d want every buyer to understand. Offering an ECG and proving that ECG is accurate are two very different things.

Wrist and band ECGs are usually “single-lead,” meaning they read your heart’s electrical signal along one path, typically by having you complete a circuit with a finger from your other hand. That can flag signs of atrial fibrillation (AFib), an irregular rhythm that raises stroke risk. Apple’s watch ECG, by comparison, went through U.S. Food and Drug Administration (FDA) clearance, which requires evidence that it works.

The Loop W’s listing mentions no FDA clearance. That doesn’t automatically mean the readings are wrong. It means nobody outside the company has shown they’re right.

So here’s our practical advice. Use the ECG as a nudge, not a verdict. If it shows something unusual, or if you feel palpitations, dizziness, or chest discomfort, see a doctor and get a proper test. And don’t let a normal reading talk you out of getting real symptoms checked.

One more detail caught our eye. The original Loop’s marketing claimed blood pressure readings and even “blood composition” analysis, the kind of claims that make health experts wince without clinical proof. The Loop W’s spec list doesn’t repeat them, which we read as a sensible step back.

The 30-Day Battery Math

Rogbid says the Loop W lasts 25 to 30 days on a 150mAh battery. That’s tiny. Many phones carry batteries over 40 times larger.

To last 30 days, the band would need to average about 0.2 milliamps of power draw. That’s possible if it takes heart-rate and SpO2 readings in intervals rather than constantly, and the efficient Bluetooth 5.3 helps. But it’s ambitious. For context, Whoop’s current band claims 14-plus days, and Fitbit Air claims about a week.

The original Loop used the same 150mAh capacity and claimed about 15 days of standby. The Loop W claims up to 60 days of standby, four times longer from the same-size battery. A newer chip could explain part of that, but until independent testing confirms it, we’d plan around two to three weeks with heavy tracking.

What It Really Costs vs Whoop, Fitbit Air, and Garmin

Price tags on screenless bands can be misleading because some brands charge monthly. Here’s the three-year picture:

Google’s Fitbit Air works without Premium, but its AI health coach sits behind the subscription. Garmin’s CIRQA band is the closest “pay once” rival from an established brand, with 10 days of battery and Garmin’s mature training metrics, but no ECG.

The Loop W wins on cost by a mile and loses on trust. Garmin and Google have years of published research behind their algorithms. Rogbid has a spec sheet. If you’re mainly curious about sleep and resting heart rate, that gap matters less. If you want heart-health alerts you can lean on, it matters a lot.

Subscription creep isn’t unique to bands, either. Samsung put a key dive feature behind a subscription on its $700 smartwatch, so the “no fees” promise is becoming a real selling point.

Who Should Buy the Rogbid Loop W?

• Buy it if you want basic sleep, heart-rate, and stress tracking on a tight budget, and you like the idea of a band you rarely charge.
• Buy it if you want to cut screen time. A band that can’t show notifications is a small, effective nudge away from your phone.
• Skip it if you need medically validated heart monitoring, or you have a known heart condition. An FDA-cleared device and your doctor are the right tools there.
• Skip it if you swim seriously. A 5 ATM rating covers handwashing, rain, and shallow swimming, not diving or water skiing.

Also consider privacy. Your heart and sleep data will live in Rogbid’s app, so check what the app collects and where it’s stored before you sign up, especially if you plan to use Family Care sharing.

The Rogbid Loop W is the cheapest credible way to try a screenless health band in 2026, and the steel case gives it a nicer feel than its price suggests. The no-subscription model is a genuine advantage, and the three-year savings against Whoop or Fitbit Air with Premium are hard to ignore.

But two big claims, the ECG and the 30-day battery, still need independent proof. Think of the Loop W as a smart, low-risk experiment in tracking your health, not a medical device on your wrist.

FAQs

Is a smartwatch better than a screenless band? 

A smartwatch adds notifications, maps, and apps, but it costs more and needs frequent charging. If you’re weighing one, the Pixel Watch 5 vs Pixel Watch 4 comparison shows what extra money buys in a mainstream watch.

Can a smartwatch detect AFib reliably? 

Some can, especially models with FDA-cleared features. Apple keeps adding health sensors, and the Apple Watch Series 12‘s biggest upgrade shows how far mainstream wrist health tracking has come. Always confirm alerts with a doctor.

Are there affordable smartwatches from Chinese brands? 

Yes, and they’re getting good. Two launched a day apart in our Oppo Watch S2 vs Vivo Watch 6 comparison, both under $300 with very different priorities.

Do screenless bands work with iPhones? 

Most do. The Loop W supports iOS 10.0 and later through its companion app, as do Whoop and Garmin CIRQA. Check whether the band syncs with Apple Health if you want all your data in one place.

The post A $60 Screenless Band With ECG? The Rogbid Loop W Makes Whoop Look Expensive, With a Big Asterisk appeared first on Memeburn.

How YouTube Video A/B Testing Works

YouTube video A/B testing is a YouTube Studio tool that lets a creator upload up to three edits of the same video or Short, serve them to different viewers at the same time, and then keep the version that holds attention best. In its creator tools post, YouTube framed the feature around finding out “which hook performs best,” which tells you where most of the experimentation is expected to happen, namely in the opening seconds that decide whether someone stays or leaves.

The workflow, as the YouTube Creators account described it, has the creator review the results, choose a winner and lock it in as the permanent version, so YouTube splits the traffic and measures the outcome but does not make the final call on its own. According to TubeBuddy’s breakdown of the announcements, each variant is scored on its share of watch time, which is the same benchmark YouTube already uses for title and thumbnail tests.

None of this is live yet. The feature is due to reach a group of select creators in 2027, and while TechCrunch’s report framed that date around Shorts, YouTube’s own announcement says the tool covers both Shorts and long-form videos. It builds on Test & Compare, which YouTube says creators have used for more than 40 million title and thumbnail experiments since its 2024 launch, and that history matters because viewers have been seeing different versions of the same video’s packaging for two years without most of them ever noticing.

The Guardrails YouTube Described to MKBHD

The two limits that shape the viewer experience most did not come from the announcement itself but from Brownlee, who said he spoke with YouTube engineers before forming his view. As Dexerto reported, YouTube plans to run each set of cuts through a semantic analysis model that scores how similar they are, and only versions that score high enough will be allowed into a test. Testing will also be restricted to the launch period of new uploads, which means a creator cannot return to a video from last year and quietly start trialling replacement edits.

Several practical questions remain open, and the walkthrough of the tool that YouTube’s Creator Insider channel published on September 25 did not settle them. PPC Land, which reviewed that demo, noted that it left out the following details:

• How long a single test runs before the creator is asked to choose a winner.
• How traffic is divided between the competing cuts while the test is live.
• What happens to the views, comments and likes that a losing cut collected.

The first of those unknowns matters more than it looks, because thumbnail tests under Test & Compare were reported to take anywhere from a few days to two weeks depending on impressions, and if cut tests run on a similar timetable, they would cover most of the period in which a typical video does its living.

Where MKBHD Is Right About Timestamps and Comments

Brownlee’s sharpest objection is that someone reading the comments has no way of knowing which version the commenter watched, and that a timestamp written by a viewer on one cut may point somewhere else entirely for a viewer on another. The arithmetic here is unforgiving, because if one cut trims 20 seconds from the opening, every timestamp after the hook shifts by those same 20 seconds, so the damage is not limited to the section that changed but runs through the rest of the video.

The timing makes this worse than it first appears. A video collects much of its early engagement in the window right after it goes live, which is also when its top comments tend to be settled, so comments written against a losing cut can stay near the top long after that cut has disappeared. Brownlee also raised the case of someone who watches one version during a test and returns later to find that a different edit has become the permanent one, which is a minor confusion for that viewer but a real problem for anyone trying to reference or quote the video afterwards.

Why Most Viewers Will Barely Notice

For the average person scrolling through their home feed, however, the change is likely to be close to invisible. The similarity check rules out swapping in a substantially different video, the hooks being tested usually differ only in their first stretch, and the test ends once the creator locks in a winner, so most viewers will see one coherent version and never learn that others existed. On Shorts in particular, where people decide within a few seconds whether to keep watching, the gap between two openings is the very thing being measured rather than something a viewer would consciously register.

That shifts the real question away from everyday viewing and towards the kinds of videos where the exact content carries weight, such as reviews, explainers and anything a viewer might later cite.

Similar to a Model Is Not the Same to a Viewer

A semantic similarity model measures whether two cuts cover the same material, which is not the same as checking whether they make the same claims. Consider a product review in which one cut includes a 15-second caveat about a known defect while another drops it to tighten the pacing, since both would almost certainly score as highly similar, yet a viewer who saw only the shorter one would leave with a different impression of the product. As TalkEsport pointed out, YouTube has announced a way to test different cuts rather than a feature designed for presenting conflicting opinions to different audiences, and there is no evidence that any creator intends to use it that way, although the system as described would not obviously prevent it either.

Brownlee expects that a viewer will probably not know whether the video in front of them is part of a test, and YouTube has not said whether there will be any on-screen indication. That sits uneasily next to YouTube’s own labels for AI-generated videos, where the platform decided that viewers deserve to know how a video was made, and a simple notice during a test would close much of the gap that Brownlee is worried about.

Better Retention Does Not Mean a Better Video

Brownlee’s broader argument is that the tool is being presented as a way to improve videos when, for nearly every creator, it is really a way to find the version with the strongest retention, and those two goals do not always line up. The concern is that openings across the platform will converge on whatever formula holds attention for the first 30 seconds, which tends to reward urgency and teasing over clarity.

The counterpoint is that creators already do this by reading retention graphs and recutting their next upload accordingly, so the tool mostly replaces guesswork with a cleaner experiment run on real audiences. For creators who already use video editing software to produce alternate openings for different platforms, YouTube is now offering to handle the measurement, which is likely to make multi-cut publishing a normal part of the workflow for channels that can afford the extra editing time.

A YouTube Video Is No Longer a Fixed Object

Seen in isolation, cut testing looks like a modest extension of thumbnail testing, but placed next to everything else YouTube has announced or is trialling, it completes a shift in what a published video actually is. PPC Land observed that once dynamic thumbnails, cut tests and dynamic brand segments are combined, the thumbnail, the opening and the sponsor message of a single video can each change after publication and differ between audiences, which the table below lays out element by element.

Brownlee’s point about a shared experience was already under pressure once thumbnails began to vary between viewers, and the difference now is that the variation reaches inside the video itself. For creators the trade is attractive, but for viewers the honest summary is that the video you watched is increasingly one version among several, and YouTube has yet to explain how anyone will be able to tell which one it was.

FAQs

Is YouTube video A/B testing available now?

No, because YouTube only announced video A/B testing at Made on YouTube on September 23, 2026, and says it will launch to select creators in 2027, with Marques Brownlee expecting a broader rollout early that year.

Will viewers know they are watching a test version of a YouTube video?

Probably not, since YouTube has not announced any on-screen label for videos in a cut test, and Marques Brownlee expects viewers will generally be unaware that they are seeing one of several versions.

Do comment timestamps break during a YouTube video A/B test?

They can, because if two cuts have openings of different lengths, every timestamp after the opening points to a different moment depending on which version the reader is watching, and that mismatch can remain in the comments after the winning cut is locked in.

Can creators A/B test old YouTube videos?

Not with video cuts, because according to Brownlee, who spoke with YouTube engineers, cut testing is restricted to the launch period of new uploads, although YouTube’s Ask Studio can separately test refreshed thumbnails on older videos if the creator allows it.

How is video A/B testing different from thumbnail A/B testing on YouTube?

Thumbnail and title tests change how a video is presented before someone clicks, while video A/B testing changes the content people actually watch, although both are scored on share of watch time and both end with the creator keeping a single winning version.

The post YouTube’s Video A/B Testing Won’t Ruin What You Watch, but It Changes What a Video Is appeared first on Memeburn.

VIVO Y500 Case + Screen Protector

Available on Amazon

Shockproof TPU protection with tempered glass, anti-scratch coverage, and an adjustable 360° ring holder.

Check Price

What Ugreen Actually Built

Ugreen calls it a 2-in-1: a protective case and a second screen. The display sits on the lower half of the back, framed by a transparent section, and it can show photos from your gallery or custom designs you make in Ugreen’s app.

TPU and PC are the two plastics used in most phone cases. TPU (thermoplastic polyurethane) is the soft, bendy part that absorbs drops, and PC (polycarbonate) is the hard shell that keeps the case’s shape.

The six colors point to E Ink’s Spectra 6 technology, the same kind of color e-paper used in some digital signs and photo frames. It reproduces bold colors well, but don’t expect the vivid look of your iPhone’s OLED screen.

The case first caught attention on social media in mid-September, with buyers on X doing double takes at the new design, and it’s now on sale across China.

How a Screen Works With No Battery

This is the clever part, and it comes down to two technologies working together.

The first is e-ink (electronic ink), the same type of screen inside a Kindle. Instead of glowing like a phone screen, it uses tiny colored particles that move to form an image. Once they’re in place, they stay there without any power. Engineers call this “bistable.” In everyday terms, the picture stays on the back of your phone for days, weeks or months, even if your iPhone dies.

UGREENの新製品なんだこれ！

Magic e-インクディスプレイケース

iPhone 18 Proシリーズ向け pic.twitter.com/g51D8IRu8J

— nakajimegame (@nakajimegame) September 16, 2026

The second is NFC (near-field communication), the short-range wireless tech behind Apple Pay. When your iPhone’s NFC chip turns on, it creates a tiny magnetic field. The case uses passive NFC to harvest that energy, so the same signal that sends the image also powers the screen just long enough to redraw it.

No battery means no charging, no extra weight from cells, and nothing to wear out after a few hundred cycles. In our view, that’s the smartest design decision here. A rear screen you have to charge separately would end up blank in a drawer within a month.

The 40-Second Reality Check

Here’s where expectations need adjusting. Sending a new picture takes about 20 seconds, and redrawing it takes another 20 seconds. That’s 40 seconds per change.

There are good reasons for the wait. NFC is slow by design, topping out at a few hundred kilobits per second, a fraction of Bluetooth or Wi-Fi. The case also only gets a trickle of power from your phone. On top of that, color e-ink particles physically move into place, and that takes time.

So this isn’t a screen you interact with. It won’t show notifications, a clock, the weather, or a camera viewfinder for selfies. Think of it like a sticker you can change whenever you want, not like a tiny second phone.

That’s fine if you know it going in. If you’re expecting the rear screen on a Xiaomi flagship, you’ll be disappointed.

The Trade-Offs Ugreen Isn’t Shouting About

A few details deserve a closer look before you order one.

The NFC button exists for a reason. The case has its own NFC antenna, and Ugreen added a side button you long-press to disconnect it. That tells us the case can get in the way of your iPhone’s own NFC. If you tap to pay with Apple Pay or use transit cards, you’ll likely want to switch the case’s antenna off first. It’s a small step, but one you’ll repeat often.

Wireless charging is a question mark. Early reports say the case doesn’t support MagSafe or wireless charging, and Ugreen hasn’t said otherwise. If you charge on a MagSafe stand every night, wait for confirmation.

Color e-ink looks muted. Spectra 6 handles bright, simple graphics well, but photos with skin tones and subtle shading tend to look flat and grainy compared to a phone screen. Bold artwork, logos and illustrations will look better than vacation photos.

It’s China-only for now. Ugreen hasn’t announced a global release. A few AliExpress sellers already list it, but you’ll pay a markup and get no local warranty.

None of these are deal-breakers at $42. They’re just worth knowing.

Who Should Actually Buy One

If you love customizing your phone and like the idea of switching your case art as often as your wallpaper, this is one of the most fun accessories of the year. It’s also a clever gift for someone with a new iPhone 18 Pro who already has everything.

If you want a functional second screen for notifications, a selfie viewfinder, or reading, skip it. You’ll be happier with a phone that has a real rear display or a dedicated e-reader.

Either way, Ugreen has proven a point. A second screen doesn’t have to cost $1,000 or need a charger. Sometimes it just needs a smarter case.

FAQs

Can you read e-books on the back of your phone?

Yes, with a different accessory. Honor’s $59 Life S3 turns the back of your phone into a Kindle with a 3.68-inch e-ink screen that clips on magnetically. It has its own 650mAh battery for about 30 hours of reading. Ugreen’s case can’t do that because it only shows static images.

Is there a full phone with an e-ink screen?

There is. E-ink phones trade a bright, fast screen for better battery life and less eye strain. The Bigme HiBreak Dual 3 is a $629 e-ink phone built around that idea. It’s a niche choice, but useful if you want fewer distractions and a screen that’s easy to read outdoors.

Is the iPhone Duo worth it just for the extra screen?

Probably not if that’s your only reason. Apple’s iPhone Duo is a foldable, so you’re paying for a tablet-sized inner display, not a rear screen. If you want more screen for less money, there are several cheaper foldable phones that rival the iPhone Duo.

Will Ugreen’s e-ink case work with other phones?

Not this model. It’s shaped for the iPhone 17 Pro and iPhone 18 Pro series, and the camera cutout and button placement won’t fit other phones. Ugreen hasn’t announced Android or base iPhone versions. Since the tech relies on standard NFC, other models are possible later.

Is it worth upgrading to the iPhone 18 Pro Max for a case like this?

No. The case also fits iPhone 17 Pro models, so you don’t need a new phone to use it. If you’re weighing an upgrade anyway, look at the iPhone 18 Pro Max’s biggest upgrades over the 17 Pro Max, like the variable aperture camera and faster A20 Pro chip.

The post Ugreen’s $42 E-Ink Case Gives Your iPhone a Second Screen That Never Needs Charging appeared first on Memeburn.

Poco X8 PRO

Dimensity 8500-Ultra power, a 6.59-inch display, and a 50MP camera built for smooth gaming and everyday performance.

 

4.5/5

• Dimensity 8500-Ultra chipset delivers flagship-class performance at 3.4GHz.
• 6.59-inch 1.5K AMOLED display with 120Hz refresh rate and 3,500-nit peak brightness.
• 6,500mAh battery pairs with blazing-fast 100W HyperCharge.

View deal

See latest price on Amazon

Huawei Mate 90 Pro Max Specs and Prices

A few terms worth unpacking: LTPO lets the screen drop its refresh rate as low as 1Hz to save battery, PWM dimming at 1440Hz reduces the flicker some people find tiring on OLED screens, and 18EV means the main camera can capture a very wide range from dark shadows to bright highlights in one shot.

The 12,000-Nit Screen, Decoded

Nits measure brightness. For context, most flagship phones peak somewhere between 2,000 and 3,000 nits across the full screen. So 12,000 sounds like a typo.

It isn’t, but there’s an asterisk. Huawei’s 12,000-nit figure is measured at 1% APL (average picture level), which means only about 1% of the screen is lit; think the glint of sun on water in an HDR video. The same 6.9-inch panel used on the Collector’s Edition and RS is rated at about 2,200 nits full screen and 6,000 nits at 20% of the screen.

2,200 nits across the whole display is the number that matters, and it’s genuinely excellent. That’s what keeps maps and messages readable in harsh midday sun. The 12,000-nit peak makes HDR highlights pop, which is real, but it’s a marketing number first.

The dual-layer (tandem) OLED design deserves credit too. Stacking two light-emitting layers lets each one work less hard, which usually means better efficiency and a longer-lasting screen. Last year’s Mate 80 Pro Max topped out at 8,000 nits, so this is a meaningful step.

One 200MP Zoom Replaces Two Cameras

Last year’s Mate 80 Pro Max carried two telephoto cameras: a 50MP macro telephoto and a 50MP super telephoto at 6.2x. The Mate 90 Pro Max swaps both for a single 200MP periscope at 4x.

At first glance, that looks like a downgrade in reach. The math says otherwise. Because 200MP is so many pixels, the phone can crop into the middle of the image to zoom further. At 8x, a 2x crop still leaves about 50MP of detail. At 12x, you’d still have roughly 22MP. That’s why Huawei calls 8x “optical-quality.”

The sensor is also large for a telephoto at 1/1.28 inch, the same size as the main camera. Huawei’s RYYB filter, which swaps green pixels for yellow ones to catch more light, should help in dim conditions. The catch is that cropping can’t beat a longer native lens in low light, because you’re using a smaller slice of the sensor.

This one-big-sensor approach is becoming the industry’s default. Oppo went even further with three 200MP cameras on the Find X10 Pro Max, and vivo is experimenting with entirely new telephoto pixel structures on the X500 Pro.

The 1-Inch Camera You Screw On

The Ruiying Z10 is the wildest part of this launch. It’s not a lens clip. It’s an entire camera with a 1-inch, 50MP RYYB sensor and a 24–240mm zoom lens (10x continuous optical zoom, f/2.0 to f/4.5) built from 19 glass elements. It screws onto the phone with a mount similar to an interchangeable camera lens and borrows the phone’s power, screen, and processing. It also records 4K video.

Early details on the Z10’s build put its weight at nearly 350 grams. Add the 235-gram Collector’s Edition and you’re holding about 585 grams. For comparison, Sony’s RX100 VII, a pocket camera with a 1-inch sensor and a 24–200mm lens, weighs about 302 grams.

Huawei built the guts of a premium compact camera and let the phone supply everything else. That’s clever. Sony tried a similar idea in 2013 with its lens-style QX cameras that paired with phones over Wi-Fi, and they fizzled. A physical mount and direct power are a smarter design, but the weight problem hasn’t gone away.

Now the catch most people will miss: the Z10 only works with the Pro Max Collector’s Edition and the Mate 90 RS. The standard Mate 90 Pro Max can’t use it. That means the cheapest way into Huawei’s 1-inch system looks like this:

Huawei also sells a CNY 19,499 ($2,910) Pro Max camera kit bundle and a cheaper CNY 1,799 ($269) SmallRig teleconverter kit. Honor is chasing the same crowd with a shotgun mic and 500mm lens mount on the Magic 9 Pro Max, so phones turning into camera rigs is clearly the 2026 trend.

The Kirin 9050 Pro and That Price Hike

Every Pro Max variant runs the Kirin 9050 Pro, Huawei’s first “LogicFolding” chip, which stacks two layers of circuitry face to face to fit more transistors in the same space. Huawei’s workaround matters because U.S. export limits block China’s chipmakers from the most advanced manufacturing tools. Huawei claims a 31% overall performance gain over the previous flagship generation, plus enough AI muscle to run a 30-billion-parameter model on the phone itself.

The price moved too. The 16GB + 512GB Pro Max costs CNY 9,999, up CNY 2,000 ($299) from the Mate 80 Pro Max’s launch price, a 25% jump that tracks rising memory costs. The iPhone18 Pro Max starts at CNY 10,999 ($1,642) in China, so Huawei still undercuts Apple, but the gap is much narrower than it was.

Which Mate 90 Pro Max Should You Buy?

• The standard 12GB Pro Max (CNY 9,499) is the value pick. You get the same screen, chip, 200MP zoom, and battery. For most people, this is the one.
• The Collector’s Edition (CNY 10,999) only makes sense if you’re serious about the Z10 module. It also adds a multispectral color sensor that Huawei says improves color accuracy by 37%.
• Skip the Z10 if you travel light. At 585 grams combined, a dedicated compact camera might be easier to carry and no less capable.

The Huawei Mate 90 Pro Max is one of the most technically ambitious phones of 2026, and all three headline specs hold up better than we expected, once you read the fine print. The screen is brilliant even without the 12,000-nit marketing, and the single 200MP zoom is a smart consolidation rather than a cut.

The Z10 is a fascinating experiment, but locking it to pricier editions makes the “1-inch camera phone” a $2,537 proposition. For most buyers, the plain Pro Max is the sweet spot.

FAQs

How does Huawei’s camera approach compare with vivo’s? 

Vivo prefers clip-on teleconverters and long native telephoto lenses rather than a full detachable camera. 

Is Honor’s top phone a better deal than the Mate 90 Pro Max? 

Honor tends to undercut Huawei on price while matching big specs. The Honor Magic 9 Pro Max vs iPhone 18 Pro Max breakdown is a good reference point for how Honor stacks up against Apple, the same rival Huawei is chasing.

Did the Mate 90 Pro Max design leak before launch? 

Yes. Early images showed a bold green finish and a ring-style camera module. The Mate 90 green design leak matched closely, and Jade Green is one of four launch colors.

Does the Mate 90 Pro Max run Google apps? 

No. It runs HarmonyOS 7, Huawei’s own system, which doesn’t include the Google Play Store. Huawei says its app store now covers more than 450,000 apps and services, mostly aimed at Chinese users.

Which other phones use the Kirin 9050 Pro? 

It debuted in the Huawei Mate XT 2 tri-fold in September and now powers the Mate 90 Pro Max, Collector’s Edition, and RS. The cheaper Mate 90 and Mate 90 Pro use the Kirin 9030 and 9035.

The post Huawei Mate 90 Pro Max: A 12,000-Nit Screen, a 200MP Zoom, and an $895 Camera You Screw On appeared first on Memeburn.