NVIDIA's $62B quarter is the least interesting part of its earnings

Jensen Said Tokens Are Money. Nobody Argued.

There's a moment in every market cycle when one company stops reporting earnings and starts reciting scripture. NVIDIA hit that moment last week.

Jensen Huang walked analysts through $62 billion in data center revenue (up 75% year-over-year) and then did something more interesting than celebrate. He handed out a lens. Compute equals revenue. Tokens are money. Performance per watt is dollars per watt. Once every hyperscaler CFO internalizes that framing, CapEx debates become circular. You're not spending on servers anymore. You're minting coin.

The market heard the number and moved on. The smarter move is to sit with the two disclosures nobody led with. First, NVIDIA put $10 billion into Anthropic. That's not a friendly gesture. That's Jensen buying a seat at the frontier model layer, not just the plumbing underneath it. Second, the Groq deal. Jensen compared it to Mellanox in the same breath he used when Mellanox was worth almost nothing. Mellanox became a $31 billion annual revenue line inside five years. Either Jensen is running the most disciplined long con in semiconductor history, or the Groq comparison is worth taking seriously.

But NVIDIA wasn't even the most structurally important call of the week. That title goes to Microsoft.

Satya Nadella reported Azure at 40% growth and an AI annual revenue run rate of $37 billion, growing 123% year-over-year. Those numbers are striking. What Amy Hood said about the business model is the actual story. Microsoft is shifting from pure per-seat pricing to seat-plus-consumption across GitHub, Dynamics, and eventually the full Microsoft 365 and Security stack. Hood was unusually plain about it: the license logic stays, but a meter gets added, just like Azure. When a CFO tells you the $600 billion-plus RPO book is being repriced upward via usage, that's the sentence you underline.

The renegotiated OpenAI deal deserves its own paragraph. Microsoft gave up royalties on model IP through 2032. In exchange, OpenAI becomes a paying Azure customer. That is not a friendship tax. That is Microsoft converting its most celebrated and most expensive partnership into a long-duration cost advantage while competitors pay retail.

Alphabet and Amazon both delivered calls that would have led the news in any quieter season. Google Cloud's backlog nearly doubling to $462 billion in a single quarter, with TPU chips now for sale at enterprise scale, means Alphabet is no longer playing catch-up in AI infrastructure. It's running a different race entirely. Amazon's Trainium chip, which Andy Jassy valued at $50 billion on a standalone basis almost in passing, is being priced by the market as a rounding error. It is not a rounding error.

And then there's Tesla, which is its own argument. Elon Musk ran that call like a man who has stopped caring about the next quarter and is building something that matures in 2030. CFO Vaibhav Taneja was visibly more cautious, pointing to negative free cash flow through the rest of 2026, tariff exposure, and one-time margin benefits. The strategic reframe is real: 1.3 million paid FSD subscribers, falling churn, rising miles per customer. Whether the car is truly just the delivery mechanism for a software product is the $800 billion question. The Street is not weighing either voice correctly right now, which means the mispricing runs in both directions.

✦

Quick Takes

Three more stories worth knowing about

Google's TPU Move Is the Sleeper Story of Earnings Season

Alphabet announced it is selling TPU chips externally to enterprise customers. Most coverage buried this under the Cloud revenue beat. It shouldn't be buried. Google built custom silicon to solve its own AI cost problem, and now that silicon is a product. The addressable market for that $462 billion Cloud backlog just expanded, because customers buying compute from Google are no longer limited to renting cycles. They can buy the hardware layer too. That is a new business, assembled without fanfare, and priced into almost no model currently running on Wall Street.

Amazon's Satellite Bet Is Bigger Than It Looks

Andy Jassy connected Amazon Leo, GlobalStar, and Apple in the same breath and called the combination 'reminiscent of AWS' in capital intensity and long-term free cash flow profile. That is not modest language from a famously careful operator. Amazon is assembling a low-orbit connectivity business in plain sight, with a hardware partner in Apple and a spectrum partner in GlobalStar. The near-term costs are real and show up in guidance. The long-term optionality is enormous and shows up nowhere in consensus estimates.

Tesla's Real Product Is Now a Subscription

Tesla's CFO Vaibhav Taneja said it plainly: the vehicle is the delivery mechanism, FSD is the product. That framing isn't new as rhetoric, but the numbers behind it are starting to validate it. Churn on paid FSD subscriptions is falling. Miles driven per subscriber is rising. Those are early but genuine signals of a software retention curve forming inside a car company's financials. The risk is that negative free cash flow through 2026 and tariff exposure on the energy segment could pressure the stock well before the Optimus and Robotaxi curves become visible to anyone outside Tesla's own projections.

✦

Overlooked Stock

SKYT

SkyWater Technology, Inc.

While the market watches NVIDIA and TSMC trade at eye-watering multiples, SkyWater Technology sits in Bloomington, Minnesota doing something neither of them do: manufacturing chips entirely on US soil, under security clearances, for customers who legally cannot buy from anyone else.

SkyWater is a pure-play US foundry serving aerospace, defense, advanced computation, and bio-health. Its customer list is effectively mandated by federal procurement rules. The CHIPS Act created a tailwind, but the more durable driver is simpler: American defense contractors need American-made silicon, full stop, and there are very few places to get it.

The company is not yet consistently profitable, and that is a real risk to hold. Revenue visibility depends heavily on government program timelines, which slip. But for investors looking for a name that sits at the exact intersection of semiconductor nationalism and AI infrastructure buildout, and is not already priced for perfection, SkyWater is worth a serious look at current levels.

The Contrarian Take

The Hyperscaler CapEx Race Ends With One Survivor, Not Five

Everyone agrees the hyperscalers are spending enormous sums on AI infrastructure. The consensus view is that this spending lifts all boats: NVIDIA, the power companies, the cooling suppliers, the data center REITs. That consensus is probably right for the next 18 months.

Here is the scenario the consensus ignores: at some point, the returns on AI compute investment become measurable and comparable across companies. When that happens, the two or three hyperscalers with the best unit economics (likely Microsoft, Amazon, and Google based on current signals) will keep spending. The others will face board-level pressure to explain why their $50 billion data center commitment is generating inferior returns.

CapEx races in tech historically end with consolidation around a small number of winners, not a permanent five-way tie. The names best positioned for that sorting are the ones already showing measurable, customer-validated AI revenue. Not ones building capacity and hoping the revenue follows.