The Infrastructure Ledger

SpaceX priced the largest IPO in recorded history today — $1.75 trillion, $75 billion raised. The S-1 it filed is also the most complete public accounting of AI compute economics ever produced. Google pays $920 million per month for compute. Anthropic pays $1.25 billion per month. On the same day, OpenAI plugged into Oracle's enterprise procurement network, and JPMorgan disclosed the enterprise AI productivity benchmark the market has been waiting for. The infrastructure economics of AI have been opaque. Today they are in the public record.

In this edition: SpaceX IPO — the S-1 as AI infrastructure balance sheet  ·  Google pays $920M/month, Anthropic pays $1.25B/month  ·  OpenAI integrates into Oracle enterprise procurement  ·  Anthropic in talks for Microsoft Maia 200 compute  ·  JPMorgan: 150,000 employees, 4 hours per week, $19.8B technology budget

The Compute Economy Goes Public

SpaceX priced the largest IPO in recorded history today — $135 per share, $1.75 trillion valuation, $75 billion raised — with Nasdaq trading under ticker SPCX opening Friday. The aerospace and satellite company going public is also the backbone of the AI compute market. Its S-1 filing contains the most complete public accounting of AI infrastructure economics ever produced, naming contracts and dollar amounts that reframe how enterprise leaders should think about the compute layer beneath every AI product they are evaluating.

These are not estimates or analyst projections. They are disclosed contract values in a public SEC filing from a company that is pricing on Nasdaq today.

The strategic context behind both contracts is the same: Colossus 1 was running at roughly 11% utilization when the contracts were signed. xAI had already migrated its primary training operations to Colossus 2, leaving Colossus 1 as an expensive, underutilized asset. Google and Anthropic — two of the most consequential AI companies in the world — are paying a combined $2.17 billion per month to rent capacity from a facility that its original builder no longer needs at full utilization. That is the supply-demand dynamic that defines AI infrastructure pricing in 2026.

The Starlink financials in the same S-1 are worth reading alongside the compute story. The satellite connectivity business generated $11.4 billion in 2025 revenue at a 63% EBITDA margin — the profitable engine that subsidizes SpaceX's aerospace operations. The AI/xAI segment generated $3.2 billion in revenue but lost $6.4 billion at the operating level in 2025. SpaceX is pricing at $1.75 trillion on the strength of Starlink margins and long-term compute contracts that convert xAI's infrastructure into predictable cash flow. At $135 per share, the implied multiple is approximately 94 times 2025 total revenue.

For enterprise leaders, the S-1 provides something that has never previously existed: a publicly filed document stating what frontier AI compute costs at scale. The organizations running cloud AI workloads on AWS, Azure, and Google Cloud are downstream of these economics. The compute scarcity that forces Google to rent capacity from a competitor sets the floor for inference pricing across the enterprise AI stack. The organizations that locked in capacity and committed to cloud pricing early hold a structural cost advantage. Those treating compute as a variable cost to be managed at deployment time — rather than a strategic input to be secured in advance — are operating in a different environment than the one their vendors built their pricing models on.

Moving Pieces

Three more developments that matter to enterprise leaders this week

OpenAI Integrates Into Oracle's Enterprise Procurement Network

OpenAI announced on June 11 that Oracle customers can apply eligible Oracle Universal Credits toward OpenAI models and Codex through Oracle Cloud Infrastructure. The integration eliminates a procurement step that has slowed enterprise adoption: organizations with existing Oracle cloud commitments can now access OpenAI capabilities within existing budget cycles and purchasing workflows, rather than initiating a separate contract cycle. Oracle has deep enterprise penetration across financial services, manufacturing, healthcare, and logistics — the same sectors that are primary buyers of enterprise AI capability. For enterprise procurement and IT leadership: if your organization has existing Oracle cloud commitments, the path to OpenAI and Codex deployment just shortened materially. Enterprise revenue accounts for more than 40% of OpenAI's total; the company is actively reducing friction in every channel where enterprise buyers already have established spend.

Anthropic in Talks to Add Microsoft's Maia 200 to Its Compute Stack

Anthropic is in early discussions with Microsoft to run Claude inference on Azure servers powered by the Maia 200 — Microsoft's second-generation custom AI accelerator, built on TSMC's 3nm process and designed exclusively for inference workloads. Microsoft claims the Maia 200 delivers more than 30% better performance per dollar compared to current GPU alternatives. Anthropic, which is simultaneously paying $1.25 billion per month to xAI for compute and has existing contracts with AWS Trainium and Google TPUs, is evaluating the Maia 200 as a potential fourth source of custom silicon. No agreement has been signed. The strategic signal is significant regardless: the frontier lab paying the highest disclosed compute bill in the industry is actively engineering its cost structure downward through supply chain diversification. For enterprise teams evaluating Claude-based products, the implication is that Anthropic's inference economics are a target, not a ceiling.

JPMorgan's $19.8B Technology Budget Sets the Enterprise AI ROI Benchmark

JPMorgan Chase set its 2026 technology budget at $19.8 billion — up $2 billion year over year, with $1.2 billion earmarked for AI investment. The operational disclosure is more instructive than the budget figure: 150,000 of JPMorgan's 315,000 employees now use the bank's internal LLM platform every week, reporting average time savings of four hours per week. At that scale, the bank's AI deployment recovers approximately 600,000 employee-hours per week. CEO Jamie Dimon acknowledged publicly that AI has displaced roles at the bank while also creating new ones — the most direct executive statement yet from a major employer on the workforce arithmetic of enterprise AI at scale. For enterprise leaders still constructing business cases for AI investment: JPMorgan has provided the external benchmark. Four hours per week, half the workforce, at an institution with 315,000 employees and a fiduciary obligation to measure every dollar. That is the defensible reference point the enterprise AI business case has needed.

the average time JPMorgan Chase employees report saving each week using the bank's internal LLM platform — across 150,000 of the bank's 315,000 staff.

At scale: 600,000 employee-hours recovered per week, at one institution. JPMorgan's internal LLM deployment has crossed the threshold from pilot to production infrastructure — used by half the workforce weekly, embedded in the workflows of analysts, advisors, engineers, and operations staff. The four-hours-per-week figure is self-reported, which typically understates actual savings rather than overstates them. For enterprise organizations still evaluating whether AI deployment at scale produces measurable returns: JPMorgan has published the number. The remaining question is not whether your organization can achieve four hours per week. It is which workflows produce the largest gains, and whether your deployment architecture is structured to capture them.

Source: AI News  ·  Emerj  ·  JPMorgan 2026 technology budget disclosure

When the Balance Sheet Goes Public

The SpaceX S-1 will be analyzed this week for Starlink subscriber growth, launch economics, and voting structure. Enterprise AI teams should read it for three numbers: $920 million per month, $1.25 billion per month, and 63%.

Those three numbers tell you something specific about the AI infrastructure your organization is building on top of. The compute layer beneath every enterprise Claude and GPT deployment is expensive, concentrated, and increasingly contracted. When Google cannot build compute fast enough to meet its own demand, the inference pricing environment for your cloud AI workloads is set accordingly. The organizations that have been treating compute as someone else's problem are looking at the same S-1 as the ones who have been managing it as a strategic input.

The JPMorgan figure — 600,000 employee-hours recovered per week across 150,000 staff — tells you the other half of the equation. The return from enterprise AI deployment at scale is real, measurable, and now publicly benchmarked by an institution that has every financial incentive to measure it accurately. The question is not whether your organization can achieve four hours per week per user. It is whether your integration architecture, change management, and use-case selection are structured to capture it.

The S-1 is the cost side. JPMorgan is the return side. Today you can read both. Build accordingly.

AK  /  Spearhead  /  Building AI systems, not tools