Chips, racks, memory and power were the day’s biggest thread by our grouping. On 12 October 2026 Microsoft, AMD, NVIDIA and Marvell set out new data-centre chips and rack designs, Counterpoint said smartphone shipments fell 20% in Q3 as memory shortages constrained supply, and the parts of FERC’s order on large loads such as data centers that it accepted, the Large Load definition and registry, took effect.
The other three threads are models and agents at work, rules and risks, and deals and capital. Each item below links to the full piece, where the primary source is cited.
Business Pill · CRITICAL PATH
A one-minute explainer of the critical path: the slowest chain sets the date. It teaches the general idea only and says nothing about any company in this roundup.
The key insight: As we read it, the day’s hardware pieces and its memory piece sit side by side: Microsoft, AMD, NVIDIA and Marvell set out chips, racks and memory expanders for data centres, while Counterpoint says continued memory shortages constrained supply for most phone makers in a quarter when shipments fell 20%.
Microsoft unveiled Maia 300, the third generation of its in-house AI chip, and said it expects the chip to deliver 40% more tokens per megawatt than leading GPU-based systems. Scott Guthrie wrote that Maia 300 will begin production deployments in 2027, and that Microsoft will scale the Maia family to multi-gigawatt deployments starting in 2027, “making Maia the primary inference system for Microsoft products and services.” Read: Microsoft’s Maia 300.
AMD introduced Helion, an open blueprint for agentic AI infrastructure based on an Open Compute Project spec-design. AMD says one example configuration, a CPU-only rack, can support up to 32,768 CPU cores, and that Helion configurations are planned to be available in November 2026. Read: AMD’s Helion rack.
NVIDIA said at the OCP Global Summit that it contributed key Vera Rubin rack and compute tray designs to OCP through its open MGX architecture, and that Meta is adopting its Spectrum-6 102.4Tb/s Ethernet switches. Read: NVIDIA’s OCP contributions.
Marvell announced the second-generation Structera X 3406 and 3506 CXL memory expanders and says the 3406 supports up to 9TB of DDR4 memory per controller. Marvell says they are expected to sample in Q1 2027. Read: Marvell’s Structera X.
Global smartphone shipments declined 20% year on year in Q3 2026, the lowest Q3 level since 2012, according to preliminary estimates Counterpoint Research published on 12 October. Counterpoint says continued memory shortages constrained supply for most phone makers. Read: Counterpoint’s Q3 estimates.
FERC, in an order issued on 9 October, rejected the core of PJM Interconnection’s plan to curtail new large loads, such as data centers, that do not bring their own new capacity. It accepted PJM’s definition of a Large Load, 50 MW or more at a single site, and a Large Load Registry, both effective 12 October. Read: FERC’s PJM order.
TTM Technologies announced a $400 million investment plan to expand capacity at four facilities in China and its plant in Penang, Malaysia. The company says the planned investments are dedicated to supporting the surge in AI infrastructure demand. Read: TTM’s capacity plan.
Tencent proposes an international offering of notes, according to its announcement to the Hong Kong stock exchange, which gives no size for the offering. Bloomberg reported, citing people familiar with the matter, that the planned offering of about $5 billion will help finance AI-related capital expenditures. The financial extract Tencent attached shows capital expenditures of RMB96.3 billion in the first six months of 2026. Read: Tencent’s notes.
SWI Group announced an €800 million capital increase with institutional investors at a €6.1 billion final valuation. It says the money will accelerate the expansion of its transatlantic AI infrastructure and compute platform. Read: SWI Group’s raise.
The day’s news of 12 October 2026, grouped into four threads. The grouping is ours; each linked story cites the primary source it was built from.
Models and Agents at Work
Meta introduced Muse Wonder and Muse Wonder Full, calling them its most capable models to date and saying they deliver state-of-the-art performance on personal and professional agentic tasks. Meta says Muse Wonder is available today through the Meta Model API and Muse Code. Read: Meta’s Muse Wonder.
Alibaba’s Qwen team announced QwenGyre, a framework for training AI agents with reinforcement learning on tasks that run for hours. Its paper says that on Qwen 3.8 2.4T it lifted the NL2RepoBench pass rate from 52.5% to 58.5% in 48 training steps, which took 75.42 hours against 134.55 hours for an Async baseline. Read: Alibaba’s QwenGyre.
Subquadratic launched SubQ Snap, a multimodal decision model that the company says handles up to 384K tokens of context. On Subquadratic’s own evaluations, Snap scores 61.7 on the Decision Index public suite against 57.9 for TypeSafe’s Jev. Read: Subquadratic’s SubQ Snap.
Codex lead Thibault Sottiaux wrote on X late on 11 October that OpenAI has raised the default context length of GPT-6.1 Sol to 500,000 tokens in Codex and ChatGPT Work, up from 272,000, which he described as “an ~ 80% increase”. Read: the GPT-6.1 Sol context change.
Sakana AI introduced Multi-Agent Self-Supervision, or MASS, in which one shared model works as a team of subagents and is then trained on its best team’s work. Its blog says MASS models won 69.9% of comparisons against the base Qwen3.6-27B after two rounds, as judged by GPT-5.5 and Claude Opus 4.8. Read: Sakana’s MASS.
A paper posted on arXiv on 28 September describes training a 4B model to rewrite the harness around an agent. On 21 task families it never saw in training, the authors report that mean revision scores rose from 0.32 to 0.62. Read: the harness-learning paper.
Igor Babuschkin said on X on 11 October that River AI’s new inference service offers its models “20% cheaper than the offerings from DeepSeek and ZAI”. DeepSeek’s pricing page shows that the rates River compares against are its off-peak rates. Read: River AI’s prices.
Linus Torvalds wrote an eight-patch Linux kernel series with Claude Code and said so in the series, dated 10 October. On his laptop, the series cut the kernel cycles per read from about 2,950 to about 725 when eight threads read the same file. Read: Torvalds’s patch series.
Similarweb put Claude at 9.6% of generative-AI website traffic in September 2026, up from 2.3% in October 2025, in an update posted on X on 12 October. On the same measure ChatGPT fell to 56.4% from 71.8%. Read: Similarweb’s traffic shares.
General Motors expects to take a new battery cell through development in “roughly half the time”, using AI models and a new $900 million mini factory in Warren, Michigan, according to a report by The Verge published on 12 October. Read: GM’s battery-cell work.
Rules and Risks
The UK government announced what its GOV.UK release calls a suite of measures to protect Britain from the risks of increasingly powerful AI. AI Minister Kanishka Narayan told the companies building the most powerful systems: “if you cannot adequately control the risks, do not release the system.” Read: the UK measures.
The UK’s AI Security Institute published its view of the emerging challenges facing frontier AI organisations and lists eight priorities for new technical best practice. It says it is “not seeking to mandate requirements for frontier developers”. Read: AISI’s eight priorities.
The European Commission said on 12 October that it will send requests for information to further investigate recent loss-of-control incidents involving AI models, after a special meeting of the AI Scientific Panel on 9 October. Read: the Commission’s requests.
Microsoft president Brad Smith told the BBC’s Big Boss Interview, published late on 11 October: “Right now, your washing machine is subject to more safety rules than AI.” He said the UK was best placed to drive global rules because of its AI Security Institute. Read: Brad Smith’s interview.
The Monetary Authority of Singapore published Version 1.1 of its SAFR white paper, a framework for checking what AI agents do before their financial actions execute; the paper’s version table dates it 9 October. The acknowledgement list adds DBS, Google and UOB. Read: MAS’s SAFR update.
NHTSA opened a file on 9 October to decide whether to investigate a petition alleging that Tesla’s driverless robotaxis exceed posted speed limits in active school zones, according to the agency’s own resume of the case. Bloomberg reported the scrutiny on 12 October. Read: the NHTSA petition.
Vice minister Li Zhong said at a State Council Information Office press conference on 10 October that China will run an action to promote employment in adapting to AI development (our translation), one of five actions for the 2026 to 2030 period. Read: China’s jobs action.
Deals and Capital
HCLTech announced a Global Premier Partnership with Anthropic and says it will set up a dedicated Anthropic Business Unit with a Center of Excellence. The release gives no financial terms. Read: HCLTech and Anthropic.
Salesforce has signed a definitive agreement to acquire Loops, which it describes as an AI-powered analytics platform, and plans to bring it into Tableau. The announcement gives no price. Read: Salesforce and Loops.
Freshworks has signed a definitive agreement to acquire Torii, which its release says tracks AI consumption to the token. The release gives no price. Read: Freshworks and Torii.
AMD announced on 28 September a definitive agreement to acquire World Labs, the AI model and research lab led by Fei-Fei Li, in an all-stock transaction valued at approximately $8.2 billion. Our piece is a catch-up read of the filing. Read: AMD and World Labs.
micro1 said it has surpassed $1 billion in annual run rate while remaining profitable, and that it priced its Series B at a $4 billion valuation. The round was led by Dream Ventures, with Microsoft and LG among the investors, according to the company. Read: micro1’s update.
Nous Research announced on 7 October that it had raised $90 million; TechCrunch reported that the round is a Series B at a $1.5 billion valuation, led by Robot Ventures. Read: Nous Research’s round.
Scarlet, a London-founded certification body for AI and software medical devices, said it has raised $50m in Series B funding led by Headline. Read: Scarlet’s Series B.
Shield AI announced on 8 October that the U.S. Navy has executed a $150 million modification to its existing agreement to advance X-BAT, and that Shield AI will invest $150 million of its own capital. The company says announced public and private commitments to X-BAT now reach $400 million. Read: Shield AI’s X-BAT funding.
OpenAI will create 200 additional jobs in Ireland, including roles to support a new Applied AI Lab at its EU headquarters, according to a release from IDA Ireland, which says OpenAI’s total commitment is now at least 450 new roles. Read: OpenAI in Dublin.
One figure from each of 12 of the day’s pieces, as each piece reports it. The figures use different units and sources and are not comparable with each other.
The Structural Read
Most of the day’s performance figures are the companies’ own, and the pieces say so: Microsoft’s 40% more tokens per megawatt is an expectation, AMD’s 32,768 cores is one example configuration, and Subquadratic’s and Meta’s scores come from their own evaluations.
The rules pieces came from public bodies on the same date: the UK government’s measures, the AI Security Institute’s eight priorities, which it says are not mandated requirements, and the European Commission’s requests for information about loss-of-control incidents.
Several deal pieces leave out a figure, and say so: Salesforce and Freshworks give no price, HCLTech gives no financial terms, and micro1 does not give the size of its Series B.
Kanishka Narayan, UK AI Minister (12 October 2026), as quoted in our piece
“if you cannot adequately control the risks, do not release the system.”
Three Implications
BUYERS OF DATA-CENTRE HARDWAREMicrosoft says Maia 300 begins production deployments in 2027, AMD says Helion configurations are planned for November 2026, and Marvell says its new CXL chips are expected to sample in Q1 2027.
DEVELOPERS OF LARGE SITES IN PJM FERC accepted PJM’s Large Load definition, 50 MW or more at a single site, and a Large Load Registry, both effective 12 October, and rejected the core of the curtailment plan.
FRONTIER AI DEVELOPERS The UK minister told them not to release systems whose risks they cannot adequately control, and the European Commission says it will send requests for information about recent loss-of-control incidents.
The framework puts it this way: “Memory is becoming the new compute. HBM growth is outpacing accelerator demand, making memory the rate-limiting factor in the AI scaling equation.”
As we read it, Monday’s pieces touch that constraint from three sides: Marvell’s CXL expanders that add memory per controller, Counterpoint’s report that memory shortages constrained phone makers’ supply, and the new rack designs from AMD and NVIDIA.
What Is Not Established
This roundup adds no new reporting. Every figure above is taken from the linked FourWeekMBA piece, each of which was built from its own primary source. The four threads and the counts in our charts are our grouping, not a finding.
We do not claim that any of these stories caused another. Counterpoint links the fall in phone shipments to memory shortages; it is our grouping, not Counterpoint’s, that puts that item next to the data-centre chips. Company figures, such as Microsoft’s tokens per megawatt or Subquadratic’s benchmark scores, are the companies’ own.
Business Engineer Framework
The AI Memory Chokepoint
A Business Engineer analysis of why high bandwidth memory, not GPUs alone, is becoming the binding constraint on AI scaling.
The day’s pieces include new chips and rack designs from Microsoft, AMD, NVIDIA and Marvell, Counterpoint’s preliminary Q3 smartphone estimates and FERC’s order of 9 October on large loads.
Alongside them, the UK government, the UK’s AI Security Institute and the European Commission each published texts on 12 October about the risks of frontier AI, and Meta, Alibaba and Subquadratic announced new models or training systems.
A note on sourcing. This roundup draws on FourWeekMBA’s news coverage of 12 October 2026, linked above; every story was built from its own primary source, which it cites and links. We re-read each piece on 12 October 2026 and took every figure here from it. The four threads and the counts in our charts are our own grouping. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
