8 min read · 10 sources

Nvidia is buying the open-source AI warehouse. Reports put the deal for Hugging Face at $12.9 billion, though TechCrunch flags that conflicting reports say a deal may not yet be signed. Either way, the logic is obvious: if OpenAI, Google, Amazon, and Anthropic are all building their own silicon, the safest play for Nvidia is to own the place every model gets uploaded, downloaded, and benchmarked.

It dropped the same week Nvidia told Wall Street that AI demand will stay supply-constrained through the end of fiscal 2028. Two days earlier, Amazon had tripled its chip order. The whole market is moving as one block, and that block answers to Jensen Huang.

Nscale’s West Virginia campus can scale to 8GW, and Microsoft has already signed for 1.35GW of it alongside Anthropic’s 460MW.

Nvidia, Amazon, and the GPU supply wall

Source: siliconangle.com ↗

Nvidia’s Q2 came in at $96 billion in revenue, up 100% year over year, with data center at $89 billion. Q3 guidance is $108 billion. CFO Colette Kress framed the multi-year picture bluntly: fiscal 2028 growth is guided at roughly 70% because that’s what supply allows, and customer forecasts point to roughly doubling the following year. Jensen Huang said capacity will come online incrementally through fiscal 2028.

For enterprise IT buyers, the practical message is that procurement is still supply-limited, not demand-limited. Lead times, allocations, and pricing are being set by Nvidia’s fab and packaging ramp rather than by competitive pressure from AMD, Intel, or the hyperscalers’ in-house silicon. If your 2027 GPU plan assumes a loosening market, plan again.

Amazon’s announcement slots into the same story. AWS is adding 2 million Nvidia GPUs across Blackwell Ultra, Rubin, and Rubin Ultra silicon, deployable in 2027 and 2028, on top of a 1 million chip order it placed five months ago. The deal extends beyond chips to networking, Nvidia’s model and software stack inside Bedrock and SageMaker, and robotics integration. AWS still ships its own Trainium and Graviton parts, but the second- and third-largest workloads by revenue at AWS are still landing on Nvidia, and the hyperscaler is willing to say so in public.

The $12.9B Hugging Face play

Source: techcrunch.com ↗

The Hugging Face acquisition is the strategic move underneath the supply story. Hugging Face is the de facto registry for open-weight models, datasets, and Spaces demos, and it is also a major hosting and inference business in its own right. Putting it under Nvidia gives the chipmaker influence over which models get promoted, which run well on Blackwell and Rubin, and how the open ecosystem lines up with its hardware roadmap.

For engineers, the consequences are practical and uncomfortable. If you ship a model on Hugging Face today, you are betting that the platform stays a neutral utility. A Nvidia-owned Hugging Face would have every reason to favor CUDA, TensorRT-LLM, and Nvidia-optimized checkpoints, and the open model ecosystem would be one corporate decision away from a quiet realignment. Watch what happens to non-Nvidia inference backends, to AMD and Apple Silicon compatibility, and to anything that competes with the parts of the stack Nvidia cares about.

The reports also conflict. The Information says talks are advanced but not finalized; other coverage is more definitive. Treat the price tag as plausible, the closing as uncertain.

The coordination tax hiding in your polyrepos

Source: differentshelf.com ↗

A long-running pattern, not a product launch, but the Different Shelf piece on repository fragmentation is a useful counterweight to all the AI excitement. The argument: in large proprietary shops, every cross-repo change triggers a market-style transaction. Owners negotiate, backlogs line up, releases get sequenced, controls re-attest. Each one is small; in aggregate, the coordination cost is a tax on shipping anything that touches more than one team’s territory.

A monorepo collapses that transaction cost by making related code discoverable and editable in one place, and the same property is what makes AI coding assistants dramatically more useful in a monorepo than across a polyrepo: the model can see more of the system. The piece is careful to note this is a business case that has to be made with measured numbers, not a religion, and that migration, tooling, and access control all cost real money. Conway’s Law points the other way too: repo boundaries frequently become team boundaries, and reorganizing the repo is often the first move in reorganizing the org chart.

Anthropic, Nscale, and 460MW of Vera Rubin

Source: datacenterdynamics.com ↗

Anthropic signed a $45 billion compute deal with Nscale for a six-year lease of 460MW at the Monarch Compute Campus in West Virginia. The site runs Nvidia Vera Rubin silicon and is targeted to come online late next year, powered in its first phase by Caterpillar G3500 natural gas generators. The campus has potential capacity of up to 8GW, with the first 2GW phase expected online in the first half of 2028. Microsoft has already signed for 1.35GW at the same site.

Stack this against Anthropic’s existing $200 billion Google agreement and the rest of its 10GW-plus of cloud commitments, and the frontier-model compute game looks less like a cloud-customer relationship and more like a power-grid and construction procurement. For an engineer, the takeaway is that frontier model capability is now gated on multi-year physical infrastructure, and the players willing to wait are the ones who will be training the next generation of models.

Z.ai open-sources GLM-5.3-Flash

Source: siliconangle.com ↗

Z.ai released GLM-5.3-Flash, previously known as Ox Alpha, as an open-weight model. The numbers to care about: 320 billion total parameters in a Mixture of Experts setup, 18 billion active per request, 1 million token input context, 131,072 token output.

The more interesting number is architectural. GLM-5.3-Flash replaces the standard softmax attention with a sparse, linear attention scheme, and Z.ai claims memory use grows linearly with prompt size rather than quadratically. Doubling a prompt roughly doubles the RAM, instead of quadrupling it. Z.ai claims a 10x cost efficiency improvement over its predecessor, with benchmark parity or wins against Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash.

Linear attention is not new, but shipping it in an open 320B model that holds up on standard evals is a meaningful data point. If you are sizing long-context inference costs, the shape of the memory curve matters as much as the parameter count, and this is one of the first public large models where you can measure the trade for yourself.

Dun & Bradstreet puts a commercial graph inside Perplexity

Source: digitalitnews.com ↗

Dun & Bradstreet is wiring its Commercial Graph into Perplexity, including the Perplexity Computer agent platform, via Model Context Protocol servers. The graph is keyed off the D-U-N-S Number and covers 650 million-plus business entities, with ownership, risk, and relationship data attached.

The integration is aimed at KYC, KYB, procurement, finance, and compliance workflows that need verified company data and an auditable trail. Dun & Bradstreet cites survey data showing 75% of organizations reporting measurable AI ROI, but only 6% saying their enterprise data is fully ready for AI at scale. The interesting pattern here is not the partnership itself but the deployment shape: an authoritative third-party data provider exposing its data through MCP, with an agent platform on the other end, and a customizable workflow layer in between. That is the template for how trusted reference data is going to ride into agentic systems over the next year.

Nutanix's $20M internal AI cluster

Source: theregister.com ↗

Nutanix built a $20 million GPU cluster internally to shift some of its Copilot and Claude usage onto its own metal, with an expected payback inside a year. Nutanix is a virtualization and hyperconverged infrastructure vendor, so a build-it-ourselves AI cluster is on-brand, but the spend and the math are the part worth noticing.

Twenty million dollars for a payback in twelve months implies a current SaaS AI bill on the order of $20M-plus, which is consistent with what other large engineering organizations are quietly spending on per-seat AI tooling. The thesis is simple: at enough seats, the per-token economics of frontier SaaS beat the per-token economics of a depreciated GPU cluster, especially for chat and coding workloads with predictable shapes. Expect more mid-to-large engineering orgs to run the same numbers through 2027.

Claude Code: leaving the Console behind

Source: support.claude.com ↗

Anthropic published a migration guide for moving Claude Code access from the Console API to Claude Enterprise. The mechanical parts are straightforward: admins provision enterprise seats, configure identity and SSO, and developers switch their login method. Local CLI session history does not need to move.

The reason to migrate is the enterprise-side control surface: centralized policies, analytics, audit logs, and hierarchical spend limits. If you are running Claude Code across more than a handful of engineers, the migration is less about features and more about getting spend and access out of personal accounts and into something finance and security can sign off on.

Palo Alto's shopping list: Okta, Datadog, and the $25B CyberArk close

Source: uk.finance.yahoo.com ↗

Palo Alto Networks CEO Nikesh Arora held acquisition talks with both Okta and Datadog before closing the $25 billion CyberArk deal. Okta is roughly a $23 billion company and talks broke down on price; Datadog carries an $80 billion-plus market cap and CEO Olivier Pomel was reportedly unreceptive. Palo Alto then picked up Chronosphere for $3.35 billion in January 2026, which contributed $338 million to its $3 billion Q4.

The strategic frame is that AI agents are about to start executing attacks, and Anthropic and OpenAI are also selling AI-powered defensive tools. Palo Alto is buying itself an identity and observability platform to compete on AI-era threat surfaces. For engineers, the consolidation story is the part that matters: the security stack is being recombined into a smaller number of very large platforms, and the boundary between identity, observability, and detection is dissolving.