Infrastructure at Scale: How Western AI Labs Are Wiring the Next Compute Era

Infrastructure at Scale: How Western AI Labs Are Wiring the Next Compute Era

The week of July 28, 2026 produced a cluster of announcements that, taken together, describe something more fundamental than any single model release: the Western AI industry is now racing to secure the physical substrate — power, land, cooling, and capital — that will determine who can run frontier workloads at all. Meta and BlackRock unveiled a $14 billion joint venture for a 1-gigawatt campus in El Paso. AMD locked in 15-year leases with Core Scientific covering 530 megawatts of capacity, with options to scale to 2.5 gigawatts. Nvidia entered negotiations to backstop a $250 billion financing package for OpenAI's proposed 10-gigawatt Ohio megasite. And Anthropic quietly shipped the most significant overhaul of the Model Context Protocol since its launch, redesigning the protocol's core to survive the kind of distributed, serverless deployments that infrastructure at this scale demands.

These are not independent stories. They are four facets of the same structural shift: the bottleneck in AI has moved from the model to the machine room.

The Meta–BlackRock Deal: Separating Capital from Operations

The El Paso announcement↗ is notable less for its headline gigawatt figure than for its ownership architecture. Funds managed by BlackRock — including Global Infrastructure Partners and HPS Investment Partners — will hold an 80% interest in the venture, while Meta retains 20%. Meta contributed land and construction-in-progress assets valued at approximately $2.3 billion; BlackRock made a cash contribution of roughly $4.9 billion; and the project carries $12.5 billion in debt financing. Total development cost is projected at $14 billion.

Meta will be the initial sole occupant, entering lease agreements with a four-year initial term and four extension options that could stretch to 20 years. The company has also provided residual value guarantees with an aggregate threshold of approximately $13 billion — a figure that underscores how seriously Meta is treating this as core infrastructure rather than a speculative bet.

"The structure separates most of the project equity from Meta's balance sheet while leaving the technology company responsible for delivery and use," as the investor release makes clear. It is a template that other hyperscalers are likely to study.

The campus is designed for 1 gigawatt of capacity, with the first power expected to come online in 2028. At peak construction, the site will employ more than 4,000 workers; the operating campus will support 300 permanent roles. Meta has also committed a $500,000 grant to local public schools for STEM and skilled-trades pathways, and BlackRock's foundation is funding a $30 million initiative to train over 12,000 electricians across Texas over three years — a telling acknowledgment that the constraint is not just land and capital, but the human workforce needed to build and maintain the infrastructure.

Why the Ownership Structure Matters

The 80/20 split is a deliberate financial engineering choice. By moving most of the equity off its own books, Meta frees capital for model development and product investment while still controlling the facility operationally. For BlackRock, the deal provides long-duration, inflation-linked infrastructure exposure — the lease includes escalators — backed by a creditworthy anchor tenant. This is the same logic that has driven airport and toll-road privatizations for decades, now applied to AI compute.

The implication for the broader industry is significant: hyperscale AI infrastructure is becoming an asset class, not just a capital expenditure. Expect more such structures as the cost of building at gigawatt scale exceeds what even the largest technology companies want to carry on their own balance sheets.

AMD and Core Scientific: Locking In the Accelerator Stack

The AMD–Core Scientific partnership↗ announced on July 28 is structured differently but reflects the same underlying logic. AMD has secured approximately 530 megawatts of critical IT capacity across five U.S. campuses — sites in Pecos and Hunt County, Texas; Muskogee, Oklahoma; Auburn, Alabama; and Dalton, Georgia — under 15-year triple-net leases with 2.5% annual escalators. Core Scientific projects more than $14 billion in base contracted revenue from the initial leases over their terms.

AMD also holds reservation rights for an additional 1.9 gigawatts of capacity through December 2028. Exercising that option in full would bring the total relationship to approximately 2.5 gigawatts — though that upper figure is an option, not a committed initial tranche.

The deal is not a conventional colocation lease. The parties are collaborating on physical infrastructure specifically designed for high-density, rack-scale platforms, including AMD's Instinct MI455X GPUs, EPYC "Venice" CPUs, the ROCm software stack, and the Helios rack-scale platform, which integrates advanced networking and HBM4 memory. As part of the agreement, AMD received market-priced warrants to purchase up to 30 million shares of Core Scientific common stock, with approximately 6.5 million vesting immediately and the remainder tied to megawatt-delivery milestones.

"For Core Scientific, the deal marks a decisive shift in its business model," as TechTimes reported↗. The company, formerly a major Bitcoin miner, now derives approximately 83% of its revenue from high-density colocation services.

AMD's Strategic Calculus

The partnership is AMD's most direct answer yet to Nvidia's infrastructure dominance. Nvidia has spent years cultivating a vertically integrated ecosystem — chips, networking, software, and increasingly, the facilities to run them. AMD is now pursuing a parallel strategy: by tying its Instinct accelerators and ROCm stack to long-term, purpose-built capacity, it creates switching costs and a deployment pipeline that is harder for customers to abandon mid-cycle.

The 15-year duration is striking. Accelerator generations turn over every two to three years; a 15-year lease implies that AMD and Core Scientific are betting on a sustained demand curve that outlasts any single chip generation. The 2.5% annual escalators and the equity warrants align both parties' incentives over that horizon.

Nvidia's $250 Billion Backstop for OpenAI's Ohio Megasite

The most structurally complex deal of the week involves Nvidia's reported negotiations↗ to provide a $250 billion financial backstop for OpenAI's lease of a 10-gigawatt campus in Pike County, Ohio — on the site of a decommissioned uranium-enrichment facility approximately 50 miles south of Columbus.

The backstop is necessary because OpenAI is currently unprofitable and lacks an investment-grade credit rating. Nvidia's guarantee is intended to reassure lenders and allow the project's developer — SoftBank's SB Energy subsidiary — to secure debt on more favorable terms. The total project cost, including hardware procurement, is expected to exceed $500 billion; Nvidia is separately discussing a financing structure to support OpenAI's chip purchases, which could total an additional $350 billion.

The first phase of the development is projected to deliver roughly 800 megawatts by 2028, with the full 10-gigawatt buildout extending well beyond that. The U.S. government is actively involved: Commerce Secretary Howard Lutnick is managing access to the power supply, and Japan has committed $33 billion in natural-gas power investment on the federal land as part of a bilateral trade agreement.

• Microsoft, Google, and Anthropic have also held discussions with the Commerce Department about the Ohio site, suggesting that the facility may ultimately serve multiple tenants rather than OpenAI exclusively.
• The deal remains non-finalized as of late July 2026; the terms are still under negotiation and there is no guarantee the financing package will close.
• The scale — 10 gigawatts — is roughly equivalent to the entire current U.S. data center industry's power consumption, compressed into a single site.

Anthropic's MCP Overhaul: Stateless by Design

While the infrastructure deals dominated headlines, Anthropic's July 28 release of the MCP 2026-07-28 specification↗ may prove more immediately consequential for developers. The update replaces the protocol's bidirectional, stateful core with a stateless request-and-response model — the most significant architectural change since MCP's launch.

In the original design, MCP servers were required to maintain persistent sessions, tracking protocol state across a continuous connection. That model works well in controlled environments but creates serious operational problems at scale: stateful servers are harder to run across serverless infrastructure, edge deployments, and Kubernetes clusters behind conventional load balancers. The new specification eliminates protocol-level session tracking entirely. Client identity, protocol version, and capabilities are now passed in the `meta` parameter of each request.

The Register's coverage↗ notes that the update also introduces several enterprise-critical features:

• OAuth 2.0 and OIDC alignment: Authorization now works with production identity systems including Okta and Microsoft Entra, allowing administrators to provision MCP connectors via central identity providers and users to inherit access based on existing group memberships.
• OAuth Mixup Attack protection: The specification now mandates inclusion and validation of an `issuer` (`iss`) parameter in authorization responses — a concrete security hardening against a known class of OAuth vulnerabilities.
• Versioned extensions framework: Capabilities that require richer behavior — interactive MCP Apps, long-running Tasks — are now handled through a formal extensions lifecycle rather than baked into the core protocol. A new deprecation policy guarantees a minimum of 12 months between deprecation and removal of features.
• MCP tunnels (research preview): Private-network tunnels allow Claude to reach MCP servers behind corporate firewalls without requiring those servers to be exposed as internet-facing endpoints.

The Security Trade-Off

Statelessness shifts complexity rather than eliminating it. Session continuity, task progress, and retry logic move into the client, an extension layer, or an external store. For enterprises, this means the authorization path, the tunnel, and the connected tools become the privileged access chain that security teams must monitor — not the protocol session itself. The OAuth hardening is a direct response to the reality that MCP servers can expose operational tools and live data, not merely return generated text. Identity and delegated permissions are now first-class concerns in the specification, not afterthoughts.

OpenAI's Scientific Agent Report: Velocity Without Validation

Rounding out the week, OpenAI published *Scientific computing in the age of agentic AI*↗, an exploratory field report examining eight case studies — primarily in life sciences — where coding agents were used to maintain, package, debug, optimize, or rewrite scientific software. Five projects used Codex exclusively; three integrated both Codex and Claude Code (Sonnet 4.5/4.6).

The headline finding is straightforward: agents accelerate well-scoped implementation work but cannot reliably determine whether their own output is scientifically correct. Researchers had to compare results against reference implementations, known benchmarks, and expected statistical behavior. Subtle numerical differences and edge cases concentrated human effort in the final stages of every project.

The full PDF↗ describes a fundamental shift in the researcher's role: from direct implementation toward orchestration and verification. Scientists are increasingly acting as product managers — defining goals, constructing tests, reviewing outputs, and deciding whether software is fit for release. Successful projects used staged, feedback-driven iterations rather than one-shot generation.

OpenAI also raised a structural warning that deserves attention: cheaper rewrites could fragment scientific software if new versions lack maintainers, ownership, or coordination with original projects. The report recommends engaging with original project maintainers as early as possible. This is not a minor operational note — it points to a genuine risk that the productivity gains from agentic coding could be offset by a proliferation of unmaintained forks that erode the reproducibility and reliability of scientific infrastructure.

The Deployment Stack, End to End

What connects these four announcements is a single observation: the deployment stack for frontier AI now extends from authorization and software interfaces all the way down to accelerators, cooling, and electricity. The MCP update addresses the software and identity layer. The AMD–Core Scientific deal addresses the accelerator and facility layer. The Meta–BlackRock venture addresses the capital and power layer. And the Nvidia–OpenAI negotiations address the financing layer that makes gigawatt-scale buildouts possible at all.

For European operators and regulators, the scale of these commitments raises a question that the EU AI Act's risk-based framework does not yet fully answer: when a single data center campus consumes a gigawatt of power and costs $14 billion to build, the infrastructure itself becomes a systemic dependency. The Act's provisions on high-risk systems and general-purpose AI models were designed with software in mind. The physical concentration of compute — and the financial structures that underpin it — may require a different analytical lens.

The week's announcements do not resolve that question. They sharpen it.

---

*Lukas Hoffmann is Neuron's Europe & Frontier Correspondent, based in Berlin.*