Data Centers

State of play.

• Compute capacity has become the decisive competitive variable. Anthropic is renegotiating cloud agreements with AWS and other hyperscalers after growth projections exploded roughly 80x in a single year, while OpenAI's earlier long-term compute commitments now appear strategically prescient .
• Hyperscaler capital commitments to AI infrastructure are operating at a scale that reshapes deal structures. Google has committed up to $40 billion to Anthropic—including 5 gigawatts of compute capacity over five years—while Amazon has committed $5 billion as part of a broader $100 billion compute agreement also providing 5 gigawatts .
• Texas grid constraints are forcing developers off the interconnection queue entirely. ERCOT interconnection delays of 5-to-10 years are pushing major developers—including Oracle with its 1.4-gigawatt West Texas facility—to build independent on-site power generation, restructuring the permitting and regulatory exposure profile for every project in the state .
• Memory semiconductor supply is constrained across the entire stack. AI data center demand has driven memory chip prices to double in Q1 2026, with further increases forecast, creating downstream cost and supply-chain exposure for consumer electronics manufacturers and their counsel .
• For counsel advising data center developers, hyperscalers, or infrastructure investors, the practical baseline is that power availability, structured capital architecture, and compute contract terms are now the three load-bearing variables in every major project—and each carries distinct legal risk that traditional real estate or technology deal frameworks do not fully address.

Where things stand.

• Compute agreements between hyperscalers and AI developers are the new critical infrastructure contracts. Google's up-to-$40 billion Anthropic commitment and Amazon's $100 billion compute agreement both include gigawatt-scale capacity commitments with performance-contingent tranches whose specific milestones remain undisclosed—creating material ambiguity around enforcement and termination rights .
• Texas is on track to surpass Virginia as the largest U.S. data center hub, with over 400 facilities in planning or construction and roughly 387 already operational—but ERCOT interconnection queue congestion is the primary constraint, with 5-to-10-year delays forcing independent power generation strategies that carry their own permitting, environmental, and utility commission exposure .
• Data center financing has moved beyond traditional equity and debt. The AI infrastructure buildout requires an estimated $5.3 trillion through 2030, with a $2.5 trillion funding gap; private credit is expected to provide roughly $800 billion of the $2.9 trillion needed between 2025 and 2028, driving adoption of SPVs, asset-backed securitizations, and hybrid equity-debt structures .
• Revenue-recognition and disclosure risk is building ahead of anticipated AI company IPOs. Anthropic's $30 billion annualized run-rate and OpenAI's $24 billion monthly recurring revenue figure are both subject to scrutiny over whether reported figures reflect paid partnerships rather than pure customer sales, with GAAP revenue materially lagging run-rate projections .
• Custom silicon and supply-chain diversification are reshaping infrastructure vendor relationships. Meta's multi-year deal for tens of millions of AWS Graviton CPU cores validates custom silicon for agentic AI workloads at enterprise scale, while Nvidia's warrant and option structure in its Corning partnership illustrates how hyperscalers are securing long-term supply commitments with embedded equity upside .
• AI infrastructure capex is driving workforce restructuring with attendant employment law exposure. Meta's planned elimination of approximately 8,000 positions—attributed directly to AI infrastructure capital demands projected to exceed $145 billion in 2026—raises WARN Act compliance, severance adequacy, and age discrimination exposure given the scale and speed of implementation .
• Enterprise AI architecture is bifurcating between hyperscaler-cloud and hybrid on-premises models. IBM's watsonx Orchestrate positioning as a multi-agent control layer targets the over 70% of enterprise data that remains on-premises, with implications for data residency, compliance, and vendor lock-in structuring in enterprise AI contracts .

Latest developments.

• Anthropic's 80x growth trajectory has forced renegotiation of major cloud and infrastructure agreements with AWS and other hyperscalers; CFO Krishna Rao is managing compute allocation, capital deployment, and revenue modeling across multiple fronts as the company prepares for potential IPO scrutiny .
• Memory chip prices doubled in Q1 2026 and are forecast to rise further in Q2, with Sony raising PS5 prices $100 and Nintendo projecting a $638 million cost increase—signaling semiconductor supply constraints that extend well beyond AI developers to consumer electronics and automotive supply chains .
• Vinson & Elkins launched its "Powering Progress" series examining the regulatory and legal architecture of data center power in Texas, with Episode 1 addressing ERCOT grid constraints and Episode 2 addressing structured capital approaches for AI infrastructure portfolios .
• Nvidia and Corning announced a multiyear partnership to expand U.S. optical connectivity manufacturing, with Corning building three new factories in North Carolina and Texas; an SEC filing reveals Nvidia holds a pre-funded warrant for 3 million Corning shares and an option to purchase 15 million additional shares in a deal estimated at approximately $500 million .
• IBM unveiled its "AI Operating Model" at Think 2026, positioning watsonx Orchestrate as a multi-agent control layer for enterprises managing hybrid on-premises and cloud AI infrastructure .
• Meta announced approximately 8,000 layoffs beginning May 20, 2026, attributed to AI infrastructure capital demands exceeding $145 billion in 2026; WARN Act and age discrimination exposure are live .
• Google committed up to $40 billion to Anthropic—$10 billion cash plus $30 billion contingent on undisclosed performance milestones—alongside 5 gigawatts of Google Cloud compute capacity over five years .
• Meta signed a multi-year deal with AWS for tens of millions of Graviton CPU cores, validating custom silicon for agentic AI workloads at enterprise scale; financial terms undisclosed .
• Anthropic's $30 billion annualized run-rate surpassed OpenAI's $24 billion figure, with both companies' disclosures drawing scrutiny over revenue composition and the gap between run-rate projections and recognized GAAP revenue ahead of anticipated IPO filings .

Active questions and open splits.

• Performance-contingent compute commitments: what triggers and what terminates? Google's $30 billion contingent tranche in its Anthropic deal is tied to undisclosed performance milestones—the same structural ambiguity likely present in Amazon's $100 billion compute agreement. How these milestones are defined, measured, and disputed will be the central drafting and enforcement question for the next generation of hyperscaler-AI developer contracts .
• Revenue recognition and IPO disclosure adequacy. The gap between Anthropic's multi-billion GAAP revenue and its $30 billion annualized run-rate—and the question of whether run-rate figures reflect paid partnerships rather than arm's-length customer sales—will face SEC scrutiny in any IPO registration. How underwriters and counsel characterize these metrics is an open and consequential question .
• Independent power generation vs. grid interconnection: which regulatory regime applies? Oracle's on-site gas generation strategy to bypass ERCOT interconnection delays sidesteps one regulatory framework but enters another—environmental permitting, air quality, and state utility commission rules for self-generation. Whether Texas regulators treat large-scale on-site generation as a utility function is unresolved and will affect every major developer pursuing the same workaround .
• Warrant and option structures in AI supply-chain partnerships: disclosure and antitrust implications. Nvidia's pre-funded warrant for 3 million Corning shares and option for 15 million additional shares—embedded in a supply commitment—raises questions about whether similar structures in other hyperscaler-vendor relationships require disclosure as material supply agreements and whether they attract antitrust scrutiny as exclusivity mechanisms .
• WARN Act and employment law exposure from AI-capex-driven layoffs. Meta's framing of 8,000 layoffs as a capital reallocation decision rather than AI displacement does not insulate the company from WARN Act, severance, or discrimination claims. Whether courts accept the "resource allocation" framing as a defense to claims of pretextual termination is an open question with implications for every large tech employer making similar announcements .
• Semiconductor supply constraints as force majeure or material adverse change triggers. Memory chip prices doubling in a single quarter—with further increases forecast through 2027—creates conditions that could trigger MAC clauses or force majeure provisions in long-term supply agreements across consumer electronics, automotive, and enterprise hardware. Whether AI-demand-driven price surges qualify as foreseeable or unforeseeable events under standard contract language is unsettled .
• Vendor lock-in and data residency in hybrid AI infrastructure contracts. IBM's positioning of watsonx Orchestrate as a multi-agent control layer for on-premises enterprise data raises the question of how enterprises negotiate exit rights, data portability, and compliance obligations when the orchestration layer—not the underlying model—holds the integration logic .

What to watch.

• Whether Anthropic or OpenAI files an IPO registration statement—triggering SEC review of run-rate revenue characterization and the adequacy of disclosures around compute commitments and partnership-derived revenue.
• Texas utility commission and environmental agency responses to large-scale on-site power generation by data center developers seeking to bypass ERCOT interconnection queues—any formal rulemaking or enforcement action will reset the permitting calculus statewide.
• Whether the Nvidia-Corning warrant structure becomes a template for other hyperscaler-vendor supply partnerships, and whether DOJ or FTC scrutiny follows as AI supply-chain consolidation accelerates.
• WARN Act litigation or regulatory action arising from Meta's May 20 layoff implementation—the outcome will signal how much insulation "AI capex reallocation" framing actually provides.
• New fabrication capacity coming online from Samsung, SK Hynix, and Micron: the timeline for relief from memory price pressure (at least one year per manufacturers' own statements) will determine whether downstream supply-chain MAC and force majeure disputes materialize in 2026 or 2027.
• Whether additional states follow Texas in becoming focal points for data center regulatory conflict, particularly as grid stress and permitting backlogs become visible in other high-growth markets.