Microsoft offers buyouts to 60,000 and OpenAI says its AI broke a 1999 limit

Yesterday the industry showed two ways of paying for artificial intelligence. Microsoft is paying with people: a voluntary exit offer to roughly 60,000 employees, more than a quarter of the company. Broadcom is paying with debt: a $63 billion structure that buys chips through a special purpose vehicle and leases them to Anthropic. In between, OpenAI published three research results it says its own models produced — including a claim that overturns an assumption computer scientists have leaned on since 1999 — while the outside experts who are supposed to check that work have not finished. Brazil's biggest bank, meanwhile, decided it would rather own its AI models than rent them.

Microsoft offers voluntary exits to about 60,000 people

Microsoft confirmed a voluntary exit offer to roughly 60,000 employees — against a headcount of 223,000 at the end of June — covering the United States, Canada, India and the United Kingdom, at vice-president level and below on selected teams.
Bloomberg reports Xbox, Windows, Copilot and Finance are taking part, while data center, sales and marketing roles are excluded; terms track standard severance. It is Microsoft's second voluntary cut of 2026, after an April retirement offer to 8,750 US employees that more than 30% accepted.

Why it mattersThe company selling the AI efficiency story is now financing it with its own payroll — and it is doing it by invitation rather than by layoff, which is cheaper in headlines and in severance.

OpenAI says its models produced three new proofs

OpenAI says internal models proved the local smoothing conjecture — posed by Sogge in 1991 — in every dimension, which also resolves the Kakeya and Stein restriction conjectures; the output is dated October 18 and the Lean formalization the next day.
It also claims a disproof of the Exponential Time Hypothesis, the 1999 assumption used to set hardness limits on satisfiability problems, plus an existence proof covering QCD and pure Yang–Mills, part of a Clay Millennium problem. Review groups led by Russell Impagliazzo and Ramamohan Paturi with Ryan Williams, and by David Gross with Arthur Jaffe and Nikita Nekrasov, are lined up but have not finished.

Why it mattersRead these as claims, not conclusions. If the ETH result survives review, a generation of complexity-theory assumptions gets rewritten; if it doesn't, this becomes the loudest AI-science correction so far — and OpenAI itself admits that choosing who verifies first creates an information asymmetry.

Wall Street puts $63 billion behind Anthropic's chips

Broadcom arranged a roughly $63 billion financing in which a new special purpose vehicle buys its AI accelerators and networking gear and leases them to Anthropic, with Broadcom giving lenders 'residual value support' — effectively a guarantee if the equipment loses value.
Bank of America, Citi and Morgan Stanley are arranging about $43.5 billion of first-lien senior debt, roughly 69% of the total, with Blackstone on the second lien. Deliveries are expected to start in 2027, the year Broadcom expects Anthropic to deploy another 5 gigawatts of TPUs. It follows a $35 billion Anthropic expansion financed by Apollo and Blackstone with a Broadcom backstop of up to $29 billion.

Why it mattersCompute is now bought the way airlines buy planes: through debt and leases, off the equity cap table, with the chipmaker partly on the hook for its own customer's risk. That works beautifully while demand holds and becomes a chain of guarantees if it doesn't.

Brazil's Itaú buys AI startup NeoSpace outright

Itaú Unibanco agreed to pay R$2.925 billion — about US$588 million at Friday's rate — for the rest of NeoSpace, a Brazilian startup founded in 2023 whose models the bank has already used for nearly two years in credit decisions, CRM and fraud prevention.
Itaú has held 15% since early 2025, when it led an $18 million round. CEO Bruno Pierobon stays, NeoSpace keeps operating separately, and the deal needs clearance from Cade and the central bank. Payment is one instalment at closing plus the rest over five years if contract conditions are met.

Why it mattersThe biggest bank in Latin America decided ownership beats a vendor contract — and that its own transaction data, not a general-purpose frontier model, is the edge.

Apple drops the 2nm M6 into the iMac and 14-inch MacBook Pro

Both machines get the same M6 that shipped in the $899 Mac mini in August: 12-core CPU, 12-core GPU and two 16-core Neural Engines that Apple rates up to 1.9x its single M5 engine, with 170GB/s memory bandwidth from 24GB up. The iMac skips the M5 entirely, jumping from M4, and Apple claims up to 5x faster on-device AI against it.
The iMac starts at $1,499 for the 10-core and $1,699 for the 12-core; the 14-inch MacBook Pro stays at $1,999 with 16GB and 1TB. Pre-orders opened today, shipping November 6 in 35 countries — designs, displays, ports and cameras are unchanged, and Thunderbolt 5 still requires the M5 Pro.

Why it mattersApple is no longer selling design changes, it is selling a neural engine — and if you want to know whether local AI is real, watch whether a $1,499 desktop is enough to run it.

The big picture

Put the day together and you get the shape of the AI build-out in 2026: the cost is moving onto balance sheets as debt and off payrolls as voluntary exits, while the scientific claims are now running ahead of the review that is supposed to confirm them. Microsoft trims a quarter of its staff list by invitation, Broadcom turns chips into a leasing business with banks on the hook, and OpenAI publishes results no outside group has finished checking. The quiet counterweight came from São Paulo: a bank that did not announce a model, a gigawatt or a proof — it just bought the thing it had been renting.

Go deeper

Sources & further reading

Topics:AI Big Tech Chips Apple