Three infrastructure stories moved on the same day — payment rails, compute capacity, and inference silicon — while Anthropic's quietly published mathematical proof signals a shift in what autonomous systems can build independently. The common thread is that AI is no longer a research project; the foundations are being poured.
Payment Networks Agree on Agent Identity Standards
On 10 September, Ant International, Mastercard, and Visa announced the Know-Your-Agent (KYA) interoperability framework, aligning three previously competing agent protocols into a single trust layer for agentic commerce. The collaboration connects Ant International's Agentic Mobile Protocol, Mastercard's Verifiable Intent, and Visa's Trusted Agent Protocol through BuildFin.ai, an open interoperability service. The announcement was made in São Paulo.
KYA does not replace each network's own verification or decisioning logic. It creates a shared onboarding and traceability layer so that an agent certified by one network does not need to re-onboard with another, reducing integration complexity and duplicative identity checks. The three organisations project that AI agents will orchestrate between $3 trillion and $5 trillion of global consumer commerce by 2030.
For operators building agentic checkout, procurement, or booking flows, the practical payoff is faster time-to-market and lower integration cost across networks. The harder question — raised by analysts covering the announcement — is what happens when agents are mis-certified, when protocols diverge under competitive pressure, or when liability shifts. Standards written for card networks were not designed with autonomous agents in mind, and the three institutions are each preserving their own decisioning processes within the framework.
Microsoft Plans to Triple Data Centre Capacity to 38 Gigawatts by 2032
Bloomberg reported on 10 September that Microsoft intends to expand its total global data centre footprint from roughly 12 gigawatts today to more than 38 gigawatts by 2032, across owned and leased facilities. AI-dedicated capacity within that footprint is projected to grow from roughly 2 gigawatts now to approximately one-third of the total, implying roughly 13 gigawatts dedicated to AI workloads alone.
The expansion follows a hard lesson in capacity management: hardware constraints in the last fiscal year forced Microsoft to turn away cloud and AI clients, restrict subscriptions, and absorb service disruptions. The company spent $145 billion in capital expenditure in fiscal 2026 and projects $175 billion for the 2026 calendar year. It also expects $50 billion in CapEx for the first quarter of fiscal 2027 alone.
For operators planning multi-year infrastructure commitments, the trajectory matters in two directions. Azure availability for AI workloads should improve substantially through 2027 and 2028 as new capacity comes online. Simultaneously, the capital intensity now required to remain competitive at this scale continues to narrow the credible field of cloud providers capable of operating at the frontier.
Positron Raises $875 Million for Memory-First Inference Silicon
Positron AI closed an $875 million two-tranche Series C round on 10 September at a $5 billion post-money valuation. The round comprised a $375 million Series C co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, and SemiAnalysis Capital, plus a follow-on $500 million Series C-1 anchored by NEA and Netscape co-founder Jim Clark.
The company's Asimov chip departs from the GPU model in a specific way: it pairs a purpose-built compute architecture with between 288 GB and 2,304 GB of commodity LPDDR5X memory per chip. The bet is that memory bandwidth and capacity, not raw arithmetic throughput, are the binding constraint when serving the largest models at the longest context windows. Asimov is scheduled to tape out on TSMC's N3P process by end of 2026, with production targeted for the second half of 2027. The target system, Titan, combines four to eight Asimov chips into a single node designed to serve models beyond 16 trillion parameters with context windows beyond 10 million tokens.
Positron is one of several new-silicon bets maturing toward production in H2 2027, alongside Etched, which shipped its first chips to Jane Street in August, and the Qualcomm-Amazon silicon partnership announced earlier this week. Each approaches the inference problem from a different constraint. Operators managing inference cost at long context should watch which architecture survives first contact with production workloads: cost-per-token at 1M-token context on today's GPU stack is the baseline to beat.
Claude Produces the First Computer-Checked Proof of Fermat's Last Theorem
Anthropic announced on 4 September that Claude had completed the first end-to-end formal proof of Fermat's Last Theorem in the Lean proof assistant, verified entirely without human intervention. The result required 13 million lines of Lean code, proved 29,500 intermediate theorems, and was confirmed by two independent checkers: Lean's own kernel using only its three standard axioms with no unverified assumptions, and nanoda, a Rust-based proof kernel that independently verified all 1,052,234 declarations. Kevin Buzzard of Imperial College London, who leads the community effort to formalise the theorem, called the result "an amazing feat of automatic formalization."
The production facts matter for operators more than the headline. Wall-clock time was 11 days; the underlying computation was roughly six billion output tokens generated by several dozen parallel agents. The first attempt failed; a mid-run integration of Columbia University's Prove2Me open-source tool was required to complete it. Trail of Bits, reviewing the claim in a technical analysis published on 9 September, noted that the result depends heavily on Kevin Buzzard's community FLT project and the Mathlib library — community-built infrastructure that Anthropic did not create but whose existence made the run possible.
The practical signal is in the workflow architecture, not the mathematics. A task that would require months of expert human time was decomposed into parallel agent runs, enriched mid-run with external tooling, and verified by independent automated checkers. That pattern — decompose, parallelise, verify independently — applies today to code audits, regulatory review, contract analysis, and drug-target synthesis. The FLT proof demonstrates the pattern holds at a level of formal rigor that few other benchmarks reach; it also shows that the supporting community infrastructure is as important as the model itself.
Taken together, today's brief describes an AI supply chain under active construction: the identity rails that agents will use to transact, the compute that will run them, the specialised silicon that will make inference at scale economical, and a capability demonstration that removes remaining doubt about what sustained autonomous reasoning can produce. Operators who are designing workflows now to match these infrastructure realities will have a measurable head start when each layer reaches production.