Chip CEOs say capital may cap the AI build-out

AI² — AI Investor Sydney · Friday 11 December, The Mint. One widely cited forecast puts cumulative AI and data-centre capital spending at about US$11 trillion by 2029. One day, six evidence-led briefings on where that capital is creating and destroying value — no panels, no pitches. Every ticket includes a year of NOOPS Pro. A$995 until 30 October → Book your seat

Two of the largest chip suppliers say money, not silicon, may now set the pace of the AI build-out, as Microsoft says it will run most of its AI inference on in-house chips and Meta puts a frontier-class model behind a subscription. Further down: Microsoft puts Copilot in front of Office while Pew finds half of US adults never use chatbots, Microsoft's president calls for rules while an open-source project writes its own, and one engineer's widely read essay argues the model that replaced them was cheaper, not better.

Suppliers now name capital as the constraint

Broadcom's Hock Tan and Intel's Lip-Bu Tan told a Singapore conference on Monday that capital, not chips, may be the biggest constraint on AI expansion, CNBC reports — Hock Tan ranking financing above shortages of semiconductors, memory or electricity, and Lip-Bu Tan adding that "in the longer run, people want to see the ROI". That is their view, and both remain bullish on demand: Hock Tan said demand for compute exceeds supply and the imbalance could persist through 2029 and possibly into 2030, and he expects the memory shortage may ease within a year, though he warned demand could slow if adoption of advanced models disappoints. CNBC links the remarks to higher bond yields raising the cost of funding data centres. Mark's view differs: in his reading, memory remains the tighter constraint. Our read is that a capital-constrained build-out favours balance-sheet-rich hyperscalers, and one of them is pulling inference onto its own silicon: Microsoft says it will mainly rely on its own chips to serve AI requests, starting with the Maia 300 accelerator unveiled on Monday, and is targeting multi-gigawatt deployments of the Maia family. Those are Microsoft's stated targets — production of Maia 300 starts in 2027 and no deployment timeline was given — and its claim of 40% more tokens per megawatt than GPUs is its own; as analyst Patrick Moorhead told CNBC, nobody has tested the chips outside Microsoft yet. Training stays with Nvidia and AMD; Evercore ISI estimates Maia will be about 10% of Microsoft's AI-chip spending by the first quarter of 2027.

A fourth lab nears the top of the index, and charges for it

Meta's Muse Wonder Full scores 56 on Artificial Analysis's Intelligence Index, fourth of the 699 models AA tracks, and Meta is putting it behind a paywall: CNBC reports US$20 or US$100 a month for Muse Wonder in the Muse app, and US$50 or US$200 for Muse Wonder Full on the Muse Code agent. Meta AI chief Alexandr Wang told CNBC that API access will cost more than the consumer plans but remain "meaningfully below" leading OpenAI and Anthropic models; AA has not yet published price or speed figures, and Meta's argument that the model generalises better to tasks it was not post-trained on is Meta's claim rather than an independent result. Per CNBC, Meta's strongest safeguards against cyber misuse come only with the paid model for now. From the other end of the market, Sakana AI's MASS has one model act as worker, workflow optimiser and judge of its own team runs, then fine-tunes on the best of them. On Sakana's results, starting from Qwen3.6-27B, the win rate against the base model on three held-out tasks, scored by external judges kept out of the loop, rose from 53.9% after one round to 69.9% after two. The public-benchmark gain, 1.2 to 1.6 times the base model, is measured per output token rather than as a raw score, and the synthetic tasks were written by GPT-5.5, so the loop is not independent of frontier models. Our read is that it points to mid-sized open models closing gaps on structured research work by learning from their own best runs.

Copilot moves to the front door; half of US adults stay away

Microsoft is folding Office into the Copilot app, The Register reports: a prompt creates an Office document, which the user edits in the matching Office application while the Copilot conversation continues, and others can prompt Copilot to revise it — for now only for Frontier programme customers. In a Message Center post over the weekend, Microsoft also said "Themes by Copilot", launched less than two years ago, will be removed from Outlook settings; it has not said why. Making the chat the entry point and Office the editing back end is the agent-first pattern. Against it sits a consumer picture that is far less settled: Pew Research Center finds 51% of American adults never use AI chatbots, The Register reports, and 67% of non-users say they are unlikely to use one in the next 12 months. Six in ten non-users cite lack of interest as a major reason, ahead of privacy and accuracy, while fewer than a third cite not knowing how — indifference, which is harder to fix with better onboarding than inability. In our view two curves are being priced as one: enterprise and developer adoption is moving fast, while consumer revenue models that assume chatbot subscriptions spread like smartphones look exposed.

A vendor asks for rules; maintainers write their own

Microsoft president Brad Smith told the BBC that AI is less regulated than milk and washing machines, The Guardian reports, and called for a legal obligation to a "victim" whose systems are breached by rogue AI agents, and for "emergency brakes" able to pause or slow AI that grows out of control. The comparison and the liability call are his; they follow Satya Nadella's call on Saturday for an AI "emergency brake", and the UK government is reportedly preparing legislation to increase oversight. One executive's interview is not policy, and who would carry a victim-liability obligation — model developer, platform operator or deploying enterprise — is the question the proposal leaves open. Meanwhile one large open-source project has written its own rule: QGIS has adopted a policy titled "No AI contributions permitted", with the enhancement proposal QEP428, opened by Nyall Dawson, accepted and merged on 7 October. One supporting comment argued that AI use from external contributors puts a "massive and unbearable pressure" on maintainers; under the previous, softer policy, around 40 of the roughly 50 pull requests tagged AI-assisted were closed without merging. The debate was not one-sided: one participant noted that every security advisory the project has received since it began accepting them was written with AI, mostly concerning its websites rather than the core, and others warned the wording would also bar AI from review and security work. In our reading the asymmetry underneath — a plausible patch costs minutes to generate and hours to review — applies wherever agent-generated code meets human review.

The case being made is cheaper, not better

An essay by the software engineer who writes as Cendyne, "Fable took my jorb", was the top post on Hacker News on Monday. It is one person's account, and it is precise about cause: "It wasn't that Fable was better at my job. As the last engineer, Fable was cheaper than my compensation." The author argues that frontier models still lack the foresight to predict the consequences of decisions over weeks, months or years, yet are good enough to change how spending decisions about people are made — and concedes that, driven step by step through a careful harness, they can port and reproduce known behaviour. The same day, two open-source projects showed "rebuild it yourself" moving from whole products down to single features: REA, published by morluto, is an MCP server that lets a coding agent inspect an app without its source code, explain how a feature works with evidence, and build a version for your own project; Doory is a self-hostable, browser-based design editor whose README pitches it as a free, open alternative to Canva. Discussing a public challenge to recreate commercial software, Mark offered a hypothesis — "People will not pay for software. People will pay for support." — which is not a finding, though it matches what Doory gives away. The exposure, in our reading, sits with seat-priced tools whose moat is a feature set rather than data, network or distribution.

Read all 10 signals →