What Moved

The OpenAI lawsuit is over. A federal jury in Oakland took under two hours on May 18 to decide that Elon Musk’s claims were filed too late under California’s statute of limitations. Musk was asking for damages exceeding $150 billion, the reversal of OpenAI’s for-profit restructuring, and the removal of Sam Altman. He got none of it.

On May 13, Kevin Warsh was confirmed as Fed Chair in a 54-45 vote. Closest in the modern era, nearly party-line, with only Senator Fetterman crossing over. Jerome Powell’s era is over. Warsh’s first FOMC meeting runs June 16-17.

On May 14, Cerebras Systems went public on Nasdaq. The company raised $5.55 billion — the largest US tech IPO since Uber in 2019. It priced at $185 per share, opened at $350, peaked at $386, and closed its first day up 68% at $311. Market cap at close: roughly $95 billion. First serious public challenger to NVIDIA in AI chips.

NVIDIA itself crossed $5.5 trillion in market cap this week, surpassing silver as the world’s second-largest asset. That number is above the GDP of every country on the planet except the United States and China. A chipmaker.

The AI Trade Goes Public

For three years, the AI trade has been a private market event. The companies shaping it (OpenAI, Anthropic, xAI, Cerebras, Cursor) existed outside the reach of most investors and outside the visibility of most finance teams. The money moving through them was invisible to standard market infrastructure.

That is ending. Quickly.

Cerebras is already public. SpaceX has filed its S-1. OpenAI is reportedly filing its confidential S-1 today, the same day this issue goes out. xAI is tracking toward a public offering. Polymarket is pricing in meaningful probability of an Anthropic IPO before year-end.

One number that makes the stakes concrete: OpenAI and Anthropic together represent 89% of annualized recurring revenue among the 34 leading private AI startups. The other 32 companies split the remaining 11%. This is functionally a duopoly. When both halves of that duopoly go public, AI gets a public market price for the first time.

And that changes the cost calculus for every enterprise in the room.

When OpenAI files its S-1, token pricing stops being a product decision and becomes a shareholder one. Right now, pricing is set by what the market will bear and what the infrastructure costs. Post-IPO, it gets set by what shareholders need margins to look like. Every enterprise AI contract signed today is pre-IPO pricing. The public market pressure coming downstream will be faster than most FinOps teams are modeling.

The window is closing. It’s worth knowing what side of it your contracts are on.

The Layer Nobody’s Tracking

On May 21, Bloomberg confirmed that Cursor’s annualized revenue hit $3 billion. One year ago, it was $100 million.

Cursor is a coding tool. Not a model, not a chip, not a cloud platform. It sits on top of models. Three thousand enterprise customers are paying more than $100,000 per year each for it. SpaceX holds a $60 billion option to acquire it, expected to close roughly 30 days after the SpaceX IPO on June 12. Cursor forecasts $6 billion in ARR by end of 2026.

None of this spend appears in standard cloud billing. There is no line item for Cursor in AWS Cost Explorer or your Azure invoice. The same is true for a growing category of AI tooling that sits above the hyperscaler layer and below most finance teams’ line of sight.

The majority of the AI spend conversation in FinOps is still anchored on API call costs and reserved instance planning. The tooling layer is being ignored. That is where the invisible spend is accumulating, and it is scaling faster than the infrastructure layer it runs on.

The FinOps Signal

New Fed Chair. Hawkish, anti-QE, rate skeptic. Warsh has been saying the same thing since 2009: that the Fed’s balance sheet expansion does more harm than good over time. Now he has the chair and a June meeting already on the calendar.

For anyone doing CapEx modeling: the rate path assumption embedded in your five-year AI infrastructure projections was built on a Powell baseline. That baseline has changed.

Infrastructure commitments made at today’s discount rates will look different if Warsh holds longer than the market expects. The commitment-to-cash-flow gap that already exists in most enterprise AI infrastructure plans gets wider in a higher-for-longer environment. It is worth running the numbers again before the June 16-17 meeting gives you the first real signal.

One more close on China: the Trump-Xi summit produced nothing for semiconductor trade. China specifically declined to approve NVIDIA H200 purchases, even after the US Commerce Department cleared them. The NVIDIA China market stays restricted. Compute costs stay elevated. Cerebras’s IPO is notable. It is not a pricing event yet. The GPU economy is still NVIDIA’s.

What to Watch

OpenAI S-1 (this week). The confidential filing reportedly happening today means the public prospectus lands in late July or August. The first leaks on token pricing and compute cost structure will tell you more about the enterprise AI cost floor than any model benchmark.

SpaceX IPO (June 12) and the Cursor close. If the listing proceeds as planned, the $60 billion Cursor acquisition follows roughly 30 days later. Watch for any pricing or contract restructuring post-close.

Warsh’s first FOMC (June 16-17). Three years of Powell Fed have been priced in. The press conference after Warsh’s first meeting is the first real data point on whether his stated positions translate to policy.

Samsung strike watch. Samsung workers are still threatening action. Memory supply is an undertracked lever on AI infrastructure costs. If it escalates, it is a trade signal for anyone holding memory names or AI infrastructure positions.

🥂 That’s the last two weeks. If someone in your world is still treating AI spend as an IT line item, this is the issue to forward. The conversation is moving to the CFO’s desk.

Talk soon,
Diana

What Moved

Two announcements dropped this week that quietly redrew the AI infrastructure map.

On May 5, The Information reported that Anthropic has committed to spending $200 billion over five years on Google’s cloud and TPU chips — a five-year agreement for 5 gigawatts of dedicated compute. Neither company has officially confirmed the dollar figure, but multiple outlets have corroborated the reporting, and Google’s backlog tells the story: it doubled to over $460 billion in a single quarter. Anthropic’s commitment alone would represent roughly 40% of that total.

One day later, on May 6, Anthropic and SpaceX announced that Anthropic will lease the entirety of Colossus 1 — SpaceX’s Memphis supercomputing facility — giving Anthropic access to 300+ megawatts of compute powered by over 220,000 NVIDIA GPUs (H100, H200, and next-generation GB200 systems). The effect was immediate: Anthropic doubled Claude Code’s rate limits for all paid tiers and removed peak-hours restrictions for Pro and Max subscribers within 24 hours of the deal closing. That’s what supply relief looks like in practice.

These two deals don’t exist in isolation. They’re the capstone on months of compute commitments that have been stacking since late 2025:

• OpenAI + AMD (October 2025): 6GW of AMD Instinct GPUs, first gigawatt deploying second half of 2026

• Meta + AMD (February 2026): 6GW, ~$60B+ over four-plus years

• Meta + NVIDIA (early 2026): ~$50B, millions of processors

• Meta + Nebius (March 2026): $27B over five years, Vera Rubin platform, delivery starting 2027

One deal notably absent from this list: the OpenAI-NVIDIA $100B framework announced in September 2025. As of this week, no contract has been signed and no money has changed hands. NVIDIA’s CFO confirmed “we still haven’t completed a definitive agreement.” It’s a useful reminder that announced capex and committed capex are two different things — and that distinction matters enormously for anyone modeling infrastructure supply.

The Infrastructure Read

Chamath Palihapitiya said it plainly on All-In this week:

“Anthropic and OpenAI’s revenue performance has nothing to do with demand. Zero. It is entirely to do with the supply constraints that exist in data centers and specifically in power. If they had infinite power, their revenues would probably be even more parabolic.”

This reframes everything. The question was never whether enterprises will pay for AI. The question is whether the infrastructure can keep up with what they’re willing to pay.

And right now, it can’t. Of the 12–16 gigawatts of AI data center capacity committed for 2026, only about 5 gigawatts is actually under construction. The rest is stuck — not because of capital or model quality, but because of transformers, switchgear, and grid infrastructure. Lead times on high-voltage transformers have stretched from 12–18 months to as long as three to five years. A $2 billion campus can sit idle waiting on a $40 million transformer order. The physical supply chain is the bottleneck, and it doesn’t care about valuation.

This creates a split: the hyperscalers (AWS, Azure, GCP) have the balance sheets to secure what exists. The labs competing for scraps face a harder road.

Meanwhile, Q1 earnings confirmed who’s winning the infrastructure race right now:

• AWS: $37.6B quarterly revenue, +28% YoY (~$150B annualized run rate)

• Azure: +40% YoY

• GCP: $20B quarterly, +63% YoY (~$80B annualized run rate) — third consecutive quarter of acceleration

GCP’s 63% is the number of the quarter. Google CEO Sundar Pichai said enterprise AI solutions became “the primary growth driver for cloud for the first time” in Q1. The consumer AI narrative stays fixated on OpenAI vs. Anthropic — but the enterprise infrastructure race may already have a quiet leader.

One data point that explains the Anthropic-Google deal beyond just capacity: Google’s TPU pricing is reportedly 40–50% cheaper than comparable NVIDIA-based solutions. When you’re committing $200 billion over five years, that spread matters.

The FinOps Signal

The hyperscaler vendor landscape is being decided right now by who can secure power — not by who has the best model.

Announced capex and committed capex are not the same thing. The OpenAI-NVIDIA situation is the case study.

A $100 billion announced commitment, backed by two of the most powerful companies in tech, hasn’t produced a signed contract in eight months. In FinOps terms: don’t model announced commitments as secured supply. Whether it’s your vendor’s capacity expansion roadmap or a headline-grabbing infrastructure deal, verify what’s actually under construction.

If your organization is mid-renewal cycle with any major cloud provider, or making AI platform decisions for 2027, the supply-side pressure your account rep isn’t advertising is real and structural. The era of elastic, abundant AI compute is over. Model it that way — and build the power and grid constraints into your cost projections.

What to Watch

• Whether the Anthropic deals officially confirm. The $200B Google figure is reported, not confirmed. Watch for Alphabet’s next investor communications for any signals.

• EWS as a hyperscaler. The Colossus deal positions Elon Web Services as a serious compute vendor. What does EWS pricing look like? What commitment structures will enterprises face? This is the new variable FinOps teams aren’t tracking yet.

• GCP’s trajectory. If 63% growth holds, the enterprise AI platform selection conversation gets complicated fast. The Google vs. everyone framing may flip sooner than expected.

• The 500-day clock. Chamath’s thesis: enterprises have roughly 500 days from the AI inflection point to prove that token spend improved operating margins. That clock is running. Watch Q2 earnings commentary for the first real signals of AI-linked margin improvement — or the absence of them.

🥂 That’s the week. If this was useful, forward it to the FinOps practitioner in your life who’s still modeling AI costs like it’s 2024. They need this more than they know.

Talk soon,
Diana

Grab a drink because this week was a masterclass in market fragility disguised as recovery. The S&P sitting around 7,135, semiconductors on an 18-day winning streak, and everyone acting like we just shook off a mild correction. Meanwhile, oil’s flirting with $100 again, the Strait of Hormuz is functionally closed, and consumer sentiment just hit the lowest reading on record. Lower than 2008. Lower than COVID. Lower than the 2022 inflation spike. But sure, let’s celebrate new all-time highs.

The disconnect isn’t subtle. Markets erased every dollar lost during the Iran war in two weeks. Tech led the charge — semiconductors up 40% since late March, NVIDIA reclaiming its $5 trillion crown, Intel posting its best single-day gain in over two decades. The rally looks impressive until you remember it’s built on a ceasefire that expires in days, peace talks that keep stalling, and oil prices that Trump himself called “peanuts” when they’re sitting at $92. He said he’d be surprised if oil wasn’t at $200. That’s not confidence. That’s a tell.

Here’s what actually happened this week: the market decided to price in optimism while ignoring every structural risk staring it in the face. Software got crushed after ServiceNow and IBM disappointed — ServiceNow down 18% in a day, dragging Microsoft, Oracle, and Palantir with it. But chips? Chips are invincible. The Philadelphia Semiconductor Index hasn’t had a down day all month. Texas Instruments beat guidance and jumped 19%. Intel beat earnings, announced a $3 billion deal with Tesla, and surged 19% in after-hours trading. The VanEck Semiconductor ETF is up 40% in 18 sessions. That’s not a rally. That’s a momentum trade on steroids.

The question no one’s asking: what happens when the AI infrastructure buildout everyone’s pricing in runs into the macro reality everyone’s ignoring? Because those two forces are on a collision course, and the impact zone is your cloud budget.

Let me connect the dots. NVIDIA’s stock is up because demand for H100s and Blackwell chips is insane. Lead times that were 52 weeks are now 26 weeks — still half a year, but better. Secondhand H100s that were going for $40K are now $12-18K because supply is finally loosening. Cloud providers are passing those savings through in spot pricing — you can get H100 instances for under $1/hour now if you’re willing to play the spot market. That’s great news for anyone running AI workloads, right?

Not so fast. Because while GPU costs are dropping, everything else in your infrastructure stack is about to get more expensive. Oil at $90+ means diesel and jet fuel prices have doubled. Airlines are cutting capacity and rerouting flights around the Middle East, adding 20-30% to shipping times and fuel costs. Cloud providers don’t eat those costs — they pass them through in the form of higher egress fees, slower cross-region replication, and tighter capacity in regions like eu-central and ap-south. Your GPU instance might be cheaper, but the data transfer to get training data in and inference results out just got 15% more expensive. And nobody’s pricing that in yet.

Here’s the other thing markets are missing: the semiconductor rally isn’t about fundamentals. It’s about positioning. Weekly inflows into global equity funds hit $48.7 billion last week, the largest inflow in 17 months, and most of that money went straight into tech. TSMC and SK Hynix both hit all-time highs this week on strong earnings. That’s real. But the XLK tech ETF is on a 16-session win streak, and the SOXX chip index has notched ten straight intraday record highs. Those aren’t organic moves — that’s momentum chasing momentum. When everyone’s long the same trade, the exit gets crowded fast.

And the exit signal is already blinking. Consumer sentiment at record lows means spending is about to contract. Gas at $4/gallon means discretionary income is getting squeezed. Airlines cutting capacity means business travel budgets are getting slashed. SaaS companies are already seeing renewal pressure — ServiceNow’s subscription revenue got “hindered by the Middle East conflict,” which is corporate-speak for “customers are delaying renewals because they don’t know what their budgets look like in six months.” If ServiceNow is feeling it, every other enterprise SaaS vendor is feeling it too. They just haven’t reported yet.

Now layer in what’s actually happening in infrastructure land. Meta just announced 10% layoffs — 8,000 people — while simultaneously ramping AI spending. Oracle laid off 3,000 employees and their stock hit an all-time high. The market reads that as “discipline.” I read it as “we’re about to see a wave of cloud cost optimization projects because companies are getting squeezed on the revenue side and can’t let OpEx keep growing.” FinOps teams are about to get very busy. Not because companies want to be more efficient. Because they have to be.

The GPU supply loosening is a double-edged sword. Yes, it’s easier to get capacity now. But it also means the scarcity premium is gone. Hyperscalers were charging whatever they wanted for H100 access when lead times were a year out. Now that supply is normalizing, pricing power shifts back to customers. That’s deflationary for cloud AI workloads, which is great if you’re buying. But it also means the revenue growth story for NVIDIA and the hyperscalers just got more complicated. You can’t keep posting 75% year-over-year datacenter revenue growth when your customers have negotiating leverage again.

And then there’s the part nobody wants to talk about: what happens when the ceasefire ends? Because it’s ending. Trump said he’s not extending it. Vance’s trip to Pakistan for peace talks got delayed because Iran didn’t respond to U.S. negotiating positions. The Strait of Hormuz is still functionally closed — tankers are turning back, gunfire reports over the weekend, the whole thing. Oil jumped 3% on Tuesday just on the fear that talks were stalling. If the ceasefire expires and fighting resumes, Brent crude goes back over $120, jet fuel doubles again, and every cloud provider with datacenters in the Middle East or heavy cross-region traffic is going to eat massive cost increases. Which, again, they pass through to you.

So here’s the operator-level reality: your GPU costs might be dropping, but your total cloud spend is about to spike. Data transfer fees are rising. Capacity constraints are tightening in key regions. Your SaaS vendors are facing renewal pressure, which means they’re going to get aggressive on upsells and price increases to hit their numbers. And if oil prices spike again, every cost center tied to logistics, shipping, or energy is going to blow up your Q3 budget.

The market’s pricing in the best-case scenario: ceasefire holds, supply normalizes, AI buildout continues, no recession. But the base case is messier. Ceasefire expires, oil stays elevated, consumer spending contracts, and companies start cutting everything that isn’t directly tied to revenue. That’s when the “AI will save us” narrative runs into the “we need to cut 20% from cloud OpEx” reality. And that’s when FinOps teams go from “nice to have” to “get in here and find savings now.”

I’m not saying don’t take advantage of cheaper GPU pricing. I’m saying understand what you’re trading off. You’re getting better access to compute, but you’re also stepping into a market where every other cost is rising, every vendor is under margin pressure, and every quarter is going to bring new surprises. The semiconductor rally is real, but it’s also crowded, momentum-driven, and vulnerable to the first sign that AI spending is slowing or that macro conditions are worsening.

The smart play right now? Lock in favorable GPU pricing while you can. Build visibility into your full cloud spend, not just compute. Watch your egress and data transfer costs like a hawk because those are the line items that are about to move. And stress-test your Q3 budget against oil at $120 and a resumed conflict, because if you’re not planning for that scenario, you’re planning to get surprised.

Markets might be hitting new highs, but the operating layer is telling a different story. Volatility is cheap right now — VIX is under 19 — which means implied vol is pricing in a smooth path forward. But realized vol is about to spike the moment something breaks. And in this environment, something always breaks.

🥂

---

Markets & Mimosas is a biweekly newsletter on markets, cloud economics, and the operating layer behind the bill.

Next edition: May 9, 2026