Capital · United States

A data-center slowdown could strengthen cloud suppliers

Credited to Rackline · Elizabeth Harding · microsoft.com · 14m ago

Microsoft and Google are choosing how scarce compute serves their own products, cloud customers and outside model developers. That gives these integrated platforms influence over the next AI expansion beyond the price of renting machines. A lab seeking more capacity negotiates with a supplier that also builds models, sells enterprise software and brings competing products to customers. The supplier has more ways to make money from the next cluster than a host selling infrastructure alone.

Restrictions on permitting or grid connections would amplify that existing allocation power if they delay usable new capacity while demand holds. The evidence is not that the election has already caused shortages. It is that working clusters are scarce now and their owners have choices. The greatest pressure falls on incremental or unreserved expansion: the buyer needs an additional commitment, rather than continued performance on an existing one. Slower demand would reverse the advantage by making capacity easier to obtain.

Microsoft has described the choice explicitly. On its Oct. 28 earnings call, Chief Executive Satya Nadella said its shared fleet lets it direct capacity toward higher-value opportunities across compute, data, Foundry and Copilot. Discussing large customers, he said Microsoft would sell capacity when the pricing was right and otherwise concentrate on building its broader enterprise business. A large buyer's willingness to spend therefore competes with Microsoft's other commercial priorities. The decision is not presented as an auction won automatically by the highest bidder.

Political influence operates through specific authorities. Texas legislators set large-load connection requirements, and the Public Utility Commission implements them; its commissioners are governor-appointed and Senate-confirmed. A security rule effective Oct. 8 covers specified new or expanded connections. Local planning powers matter too: Manatee County commissioners approved a 12-month pause on new data-center-related applications in the county's unincorporated area. Both measures predate the election. Future officials can change policy through those institutions, but election results impose no national construction freeze.

The economic effect depends on whether those decisions delay working clusters. Fewer timely permits or grid connections mean less usable new supply; if buyers still need it, existing owners choose which additional workloads to serve, affecting rivals' expansion costs and schedules. Electricity, cooling, networking and chips must arrive together. Anthropic identifies transformers and turbines among its infrastructure risks. These constraints also reach hyperscalers' own future projects. Their existing powered fleets and their construction pipelines are different investments, and a connection queue measures neither available GPUs nor completed clusters.

Google also treats allocation as a choice across businesses, rather than just a sale of machine time. On Alphabet's Oct. 28 call, Chief Executive Sundar Pichai described long-range investment-return planning and said its models underpin Search, YouTube, Cloud and other businesses. Microsoft likewise develops its own MAI models for its software. Both companies can weigh a lab's infrastructure purchase against uses inside their own platforms. That gives them alternatives when negotiating expansion, without requiring them to reserve all capacity for themselves.

Software distribution supplies a separate source of influence. Microsoft says 90% of Copilot queries use automatic routing across models, optimizing for quality, speed and cost. Its Foundry platform also connects outside models to enterprise customers. Those product choices influence which models get used, independently of whether computing capacity is scarce. Together, the two capabilities let a cloud discuss both the resources a lab needs and the customers it hopes to reach in one commercial relationship.

Capital-backed partnerships make that relationship tangible. In Nov. 2025, Microsoft announced an investment commitment of up to $5 billion in Anthropic alongside Anthropic's $30 billion Azure purchase commitment and access for Claude through Foundry and Copilot products. Those are the historical announcement's terms, not today's total commitments. The package shows how a cloud can help finance a lab, supply its compute and distribute its models. For Anthropic, that supports more model development and customer deployments while giving Microsoft a negotiating role in the scale and timing of its Azure expansion.

Reservations turn access into a financing decision even without an investment deal. Google's published rules say provisioned reservations hold resources and require payment whether they are used or not. An allocation quota, by contrast, does not guarantee that resources are available. A startup planning a large deployment must therefore consider securing capacity before its usage arrives, finding another supplier or changing the workload. That can consume financial flexibility earlier in its growth. For a startup with uncertain usage, reserving early shifts more of the risk of underused machines from the provider to the buyer.

Labs that already secured long reservations can gain against entrants too. OpenAI's reserved-capacity orders with CoreWeave establish a direct route outside the biggest integrated clouds, while Anthropic's filing documents multiyear cloud spending commitments. Once reserved resources are delivered and fit the workload, a lab can add training runs or serve more customers while an entrant without that access must first negotiate an allocation.

Specialists can capture value from the hyperscalers themselves. Nebius agreed to provide Microsoft dedicated capacity from Vineland, New Jersey. In its Sept. 2025 announcement, Nebius expected to finance the buildout through customer cash flow and debt secured against the agreement, with terms improved by Microsoft's credit quality. That financing route uses an anchor customer's commitment to support an independent builder's expansion. Capacity dedicated to Microsoft is not automatically available to another lab, but the owner of deliverable power has bargaining power upstream.

Off-cloud hardware makes the counter-case stronger. Google says it now supplies complete TPU systems for customer data centers and is expanding availability through other providers. Alphabet expects a meaningful system-sales revenue increase in 2027. If independent powered sites can combine those systems with competitive service, labs gain an alternative to buying the entire cloud package. Google also says third-party bridging capacity will pressure Cloud margins. That is a concrete reason integrated platforms cannot assume they capture all the value of scarce supply.

Efficiency can shorten the advantage before another building opens. Microsoft reported more than 50% higher quarterly throughput for Copilot workloads and said broader engineering improvements enabled additional customer allocations. The throughput figure describes those workloads, not a fleetwide increase in availability. Microsoft also expects delayed third-party capacity during its fiscal second half. Actual deliveries, combined with gains that exceed usage growth, would give buyers additional options and weaken the leverage attributed to shortages. The advantage lasts only while demand keeps absorbing the compute released by engineering gains and new deliveries.

Current disclosures still favor suppliers with usable capacity. Microsoft says demand exceeds available supply; Alphabet describes a constrained environment despite its own chips and outside purchases. Amazon offers a narrower pricing signal: in its own account of third-quarter earnings remarks, Andy Jassy said short-term prior-generation Trainium capacity sold at a premium to the first half of 2026. It supplied no premium percentage or matched price series.

When deliverable expansion is scarce, Microsoft and Google can negotiate capacity alongside financing and distribution, giving them influence over how quickly an AI business grows. A well-funded entrant still needs a supplier willing and able to deliver. Independent capacity that is usable and delivered on time gives the entrant a way to expand without relying on the integrated platform's next offer.

CapitalUnited States

Applied Compute raises $350 million at $3.25 billion valuation

Credited to Rackline · Daniel Reeves · appliedcompute.network9h ago

Useful

Applied Compute has raised $350 million at a $3.25 billion valuation in a round led by Elad Gil, the company said on Nov. 2. Nvidia, Blackstone, Kleiner Perkins and Lux Capital also invested.

The company said its annualized revenue run rate had reached $100 million less than 18 months after its founding. It plans to spend the money on compute capacity, expand its team and speed up enterprise deployments, without saying how the funds will be split.

A run rate projects recent sales forward rather than reporting a year of earned revenue. Applied Compute did not give the period or method behind its figure.

Latham & Watkins, which advised Applied Compute on the deal, gave the same amount, valuation and investor list in a Nov. 3 announcement. Neither announcement said whether the valuation was before or after the new money, what security investors bought, or how much each investor put in.

Applied Compute builds and runs models tailored to customers' own work. Its AC2 platform combines training, inference and continual improvement, the company said, and customers have completed more than 1,700 training runs on it since July.

Its inference service now handles more than 1 trillion tokens a day, the company said. It added that 40% of customers use it for both training and inference.

The announcements disclose no supply, procurement or customer agreements attached to the investments.

Applied Compute named Microsoft, Shopify, Mercor, Harvey, DoorDash, Cognition, Clay and Decagon as customers. It said the number of enterprises it serves had more than doubled since its previous fundraise, without giving a total or any customer-level revenue.

The company also described two customer results. It said its work on Harvey's Review Table product improved answer and citation quality at less than half the inference cost of a frontier alternative. It said a custom router it built for Clay cut production inference costs by up to 47% without sacrificing quality. Both are company claims.

Source · appliedcompute.network

Front

Capital · United States

Anthropic pushes ahead with pre-Thanksgiving IPO despite tech selloff, accepts lower valuation

Credited to The Decoder · the-decoder.com · 11h ago

Anthropic is pressing ahead with its IPO despite the tech selloff, accepting a lower valuation. The Claude developer wants to go public before Thanksgiving and raise $80–100 billion at a $1.25–1.75 trillion valuation, according to investors cited by the Financial Times. That tops the current $965 billion but falls short of the $2 trillion the most optimistic investors had hoped for. Annualized reported revenue now sits at about $81 billion, up from roughly $65 billion at the end of July.

Rising borrowing costs and heavy AI spending plans triggered a tech selloff this week. Democrats winning the US midterms accelerated it because they had promised to slow data center construction, including other measures. The Nasdaq 100 dropped 6.4 percent on Thursday but recovered somewhat on Friday.

OpenAI also cut its funding-round target from $1.4 trillion to $1.25 trillion and recently raised only about a third of the $30 billion it originally sought. Ad Ad

AI News Without the Hype – Curated by Humans