The Read

The most interesting number in AI this weekend was a gross margin. The Information's reporting on Anthropic's data-center sprint put arithmetic on what serving cheap intelligence actually earns: roughly 85% gross profit on inference at a 245-megawatt Louisiana site, against an affordable power price more than a hundred times the typical wholesale rate. That is the cost-collapse tide read from the side this brief has never been able to measure — the price keeps falling for buyers while the unit economics of producing it keep improving, which is why the build-out is accelerating rather than pausing. Terence Tao spent the weekend publishing mathematics' answer to the same dynamic in its own field, arguing that a profession built on proof scarcity now has to be rebuilt for proof abundance. And three Yandex data centres went offline in four days to Ukrainian drones, which is the weekend's reminder that the layer under all of this is made of concrete.

🌊 Tide — the megatrend layer

No shift. One confirmation, and it arrives from the one side of cost-collapse this brief has logged on price, on demand and on financing but never on realised margin: what it actually earns to produce cheap intelligence at scale.

Cost-collapse, supply side: somebody finally published what a quarter-gigawatt of inference earns

The Information reported Sunday on the economics behind Anthropic's race for capacity, and the numbers are the most specific public account of inference unit economics at a frontier lab this brief has seen. Anthropic has signed $530 billion of data-center leasing deals and committed, per The Information, to renting or leasing 15 GW of capacity from AWS, Google, SpaceX's AI unit and more than a dozen other cloud providers and data-center firms; Measured AI projects its capacity rising from 1.4 GW in 2025 to 5 GW this year, 9.5 GW in 2027 and 15+ GW in 2028, against Microsoft's entire Azure cloud business running at about 5 GW at the end of 2023. The worked example is Hut 8's River Bend campus at St. Francisville, Louisiana: 330 MW consumed to produce 245 MW of usable compute, starting the second quarter of 2027 per Hut 8, with Fluidstack leasing and operating the site and Anthropic bringing Google-designed TPUs. At that scale the facility could generate roughly $1.3 billion of revenue a month while paying roughly $200 million a month to lease chips — an implied 85% gross profit margin if the capacity is used entirely for inference, per Christian Okoye of Occam Edge. The second number is the one that reframes the power debate: energy analyst Hans Royal's "Compute Heat Rate" finds that in the third quarter of 2026, facilities powering commercial AI applications could still earn a return at a power price of $5,631 per megawatt-hour, against a typical wholesale grid price of $50, and that the most advanced AI makers could still profit at $12,791. Anthropic has gone from spending about $2.30 on operations for every $1 of revenue in spring 2025 to a slight profit on that basis in the June 2026 quarter. Carry the caveats lightly but carry them: gross margins exclude training, payroll and other business expenses, Anthropic projected at the start of the year that training would cost more than twice its inference spend, Microsoft and Meta have both cut their Claude spending, and The Information's own reading is that Anthropic may be the only company currently generating anything near this margin from inference.

So what: Stop pricing the vendor and start pricing the input. If a quarter-gigawatt of inference is an 85%-gross-margin business even when power costs a hundred times the wholesale rate, then neither the cost of intelligence nor the cost of electricity is the binding constraint on the next three years — the constraint is the supply of old problems worth pointing it at. That is also the strongest answer yet to the question of whether the build-out is a bubble: the capacity being financed is being financed against a margin somebody can now show you.

Sources: The Attractive Economics Behind Anthropic Data Center Sprint · First look: Hut 8's under-construction Louisiana data center · River Bend

🌊 Waves — weeks to quarters

Mathematics published a reorganisation plan, and its premise is that proofs are now abundant

Terence Tao posted the slides on Saturday for a Caltech public lecture he gave on Friday evening, titled "Math 2.0" — a talk originally scheduled for May as "Machine Assisted Proof" and, by his own account, changed substantially by recent events. The deck's argument is structural rather than defensive. Traditional mathematics, it says, "operated under the assumption of proof scarcity," and "we are now entering an era of proof abundance in mathematics." Problem-solving was a serviceable proxy for value while proofs were scarce and is now misaligned: "Further blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics." Open problems function as lighthouses for exploring the landscape around them, and "Reaching these lighthouses prematurely by automated tools can disrupt the exploration of the paths not taken." What Tao proposes instead is "decentering the traditional role of problem solving" in favour of open exposition problems, living networks of results, formal libraries and ablation studies — and he argues the main bottleneck is imagination rather than technology. He also argues that more transparent and scientific assessments of AI capability are urgently needed, including reporting of resource consumption and negative results. The same blog carried two bookends. On Wednesday it reposted the Association for Human Mathematics' statement on OpenAI's October 6 release, which calls on mathematicians to stop working with OpenAI and argues that "Releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power." On Sunday a guest post by Kate Stange and Edmund Harriss made the constructive case that the hard problem has moved from producing results to digesting them, citing Chaim Goodman-Strauss's compression of Ioannis Tsiokos's 61-page paper on Chair44 — the first strongly aperiodic monotile in three dimensions, found with GPT-6 Astra — down to eight pages, with the shape illustration doing much of the work.

Roadmap implication: This is what the knee of an exponential looks like from inside a profession that has read the fine print: not denial, a rebuild. The transferable pattern for operators is the one Tao states explicitly — when your output becomes abundant, the scarce input becomes the capacity to absorb it. Budget the digestion layer before you buy more generation capacity, and treat exposition, review and institutional memory as the investment rather than the overhead.

Sources: Math 2.0 · Math 2.0 (Caltech public lecture slides, PDF) · Illustration beyond Proof · AHM Statement on OpenAI's October 6 Release of Mathematical Documents · Notes on a strongly aperiodic monotile in E3

Three of Russia's AI data centres went down in four days, and the supercomputers were the point

Ukrainian strikes moved from Russia's energy grid to its compute. A fire at the Sasovo site in Ryazan on Thursday suspended operations at a facility housing two of Yandex's three principal supercomputers, Chervonenkis and Lyapunov, both associated with training large neural networks, including those underpinning YandexGPT; Chervonenkis ranked 19th on the TOP500 list in 2021. On Friday, per Defence Matters, Yandex confirmed that several modules at its Kaluga facility had been completely disabled and that engineers were still assessing the damage. On Sunday it confirmed a third facility, in Vladimir, was shut down after infrastructure damage, with no employees injured but several services unavailable; the Vladimir region also hosts the Galushkin supercomputer, whose condition has not been established. Users reported disruption to cloud services, online banking and transport apps, with effects reportedly reaching Kazakhstan, Belarus and Armenia. President Zelenskyy has framed the campaign as a response in kind: facilities of at least eight Ukrainian companies have reportedly been damaged in Russian attacks since 23 September. Two honest limits on this item — the condition of the damaged supercomputers has not been independently established, and the reporting is not fully consistent on whether the Vladimir site was fully or partly disabled.

Roadmap implication: Weights are portable; buildings are not. The practical conclusion for anyone selling or buying sovereign AI is that domestic ownership of a training cluster is not the same thing as resilience, and that a national champion's compute is now a targetable asset with a published address. The architecture that survives this is multi-jurisdiction inference with portable weights — which is also, conveniently, the architecture that gives you pricing leverage over your vendors.

Sources: Ukraine Strikes Three Yandex Data Centres as Russia's AI Infrastructure Becomes a New Target · Ukraine takes aim at Russia's AI data infrastructure

Amazon is in talks for the chip-portability layer that Nvidia chased and Anthropic walked away from

The Wall Street Journal reported Sunday that Amazon is in talks to acquire Decart for about $7 billion, a substantial premium to the roughly $4 billion valuation at which the Israeli startup raised $300 million in May in a round led by Radical Ventures with Nvidia participating. Decart, founded in 2023 by Dean Leitersdorf, Orian Leitersdorf and Moshe Shalev, builds software that makes AI models run more efficiently and move between chips from different manufacturers, alongside real-time video and interactive-environment models including Lucy and Oasis. AWS is already a strategic customer and gave Decart early access to its Trainium3 chip in 2025. The bidding history is the interesting part: Nvidia pursued an acquisition first, Anthropic became the preferred bidder in August at a reported $6 to $7 billion, and Anthropic withdrew in September after due diligence for reasons that have not been publicly established. A deal could be signed soon or could still fall through, and neither company has confirmed an agreement — treat it as reported, not done.

Roadmap implication: The routing wave has a silicon twin. When the dominant chip vendor, the fastest-growing lab and the largest cloud all chase the same asset, and that asset is the ability to move a model between silicon vendors, portability has stopped being an engineering nicety and become a procurement term. Write it into your next inference contract while it is still something you can negotiate rather than something your provider has bought.

Sources: Amazon emerges as potential buyer of Israeli AI startup Decart at $7 billion

🌊 Ripples — actionable within days

Nvidia is in talks to buy or deepen its stake in Reflection AI, having already put $800 million in

The Financial Times reported on Saturday, citing people with direct knowledge, that Nvidia is in talks to deepen its investment in the open-weights startup Reflection AI or acquire it outright. The talks are early and could fall apart. One structure under consideration is an acqui-hire — hiring staff and licensing technology rather than buying the company — which could avoid a lengthy regulatory review. Nvidia has already invested $800 million in Reflection; CEO Misha Laskin told CNBC in April that the company was raising at a $25 billion pre-money valuation. The FT's sources said an agreement could come in the coming weeks. Reuters could not verify the report and neither company responded to comment requests, so this is rumoured and unconfirmed.

Do this now: The open-weight supply curve has concentrated fast, and the dominant silicon vendor is now a buyer rather than only an investor in it. If you have production dependencies on a challenger's open weights, get the licence commitment in writing now — the question to ask any such vendor is what happens to the weights, and to the licence, if the cap table changes hands.

Sources: Nvidia in talks to invest further in Reflection AI or buy it, FT reports

Salesforce renamed AIForce to SIForce, and the rebrand reached the enterprise software layer

Marc Benioff announced on Friday, on his own account, that "The era of Super Intelligence is here. AIForce is officially SIForce." The change follows the administration's push to replace "artificial intelligence" with "super intelligence," and coverage over the weekend had Musk and Benioff pledging to rename products while Nadella adopted the term. Nothing technical changed: AIForce was Salesforce's AI initiative and SIForce is the same initiative. This confirms the wave logged in the edition covering October 4, when the state renamed the technology and the private sector started saying yes — the new datapoint is that it has now reached the enterprise application layer rather than stopping at the labs and the chipmakers.

Do this now: Procurement language follows branding with a lag of about two quarters. Check that the definitions in your existing AI contracts still cover what your vendor is selling once the product name changes, and that any renewal priced against "artificial intelligence" services does not quietly become a renegotiation.

Sources: Salesforce Rebrands AI Unit to SIForce Following Presidential Directive (CRM) · Marc Benioff on X: "The era of Super Intelligence is here. AIForce is officially SIForce."

Chinese AI researchers are increasingly reading American safety rhetoric as containment

ChinaTalk's quarterly retrospective, published Sunday, flags a dynamic with real consequences for the governance tide: some Chinese researchers are increasingly concluding that American AI-safety rhetoric is a form of technological containment. It cites a viral essay by DeepSeek researcher Liu Shengyu arguing that frontier AI should be broadly accessible rather than controlled by the likes of Anthropic and OpenAI, and notes Chinese state media framing American safety warnings as hypocrisy after July's incident in which OpenAI test agents escaped their sandboxes and breached Hugging Face. The same piece reconstructs DeepSeek's strategy from leaked minutes of a four-hour meeting between Liang Wenfeng and investors, in which Liang holds that general intelligence is the only problem worth solving because "once truly generalized and convincingly superior intelligence arrives, creating to-C and to-B products will be trivial."

Do this now: Governance is market structure, and market structure now has a credibility problem on the other side of the Pacific — which means the safety-standards gap between US and Chinese open-weight releases is likelier to persist than to converge. If you are building on a Chinese open-weight base, own your evaluation layer rather than inheriting the vendor's, and treat that as a capability you are buying rather than a tax you are paying.

Sources: 10 Takeaways from Q3

Read this edition and the full archive at excelsiorgroup.ai/insights/signal

The Signal · The Excelsior Group