NVIDIA beat every number that mattered, and the stock still slipped. What that tells you about where the AI money is heading is far more interesting than the chipmaker itself.
Multi-trillion dollar chipmaker Nvidia did the thing it always does on 20 May, which is to make the analysts paid to forecast it look slightly silly for trying.
Revenue landed at US$81.6 billion, up 85% on the year, with second-quarter guidance of $91 billion sitting ~$4 billion north of what the Street had pencilled in, and gross margins parked at a frankly indecent 75%.
For good measure, the board lifted the quarterly dividend 25x and authorised another $80 billion of buybacks, while free cash flow hit a record $49 billion in a single quarter.
Then the stock fell.
Not dramatically - just the muted after-hours dip of a market that had already moved on - and that small, almost petulant slide is the real story of this earnings cycle.
Wall Street has spent the past few quarters dragging its own price-in above the company's official guidance, which means an outperformance now requires NVIDIA to clear a threshold it never actually set for itself.
Daniel Newman of Futurum Group framed it succinctly on the night.
"Wall Street has pulled itself above the company's own guide for the first time in this cycle, which raises the bar for what counts as a beat," Futurum Group chief executive Daniel Newman said.
Capital.com's Kyle Rodda was blunter still.
"This was a garden-variety beat - a better-than-expected top and bottom line with guidance above the Street estimate - and one that was well telegraphed following the very strong results from AI hyperscalers earlier in the earnings season," Capital.com senior market analyst Kyle Rodda said.
When the most consequential hardware name in the cycle delivers a flawless print and the share price shrugs, the question stops being whether appetite is genuine and becomes where the operating leverage migrates next.
Under the numbers
The engine beneath all this is hyperscaler capital deployment, and the figures have left the realm of the comprehensible.
Microsoft, Amazon, Alphabet and Meta are guiding to combined capex of ~US$725 billion in 2026 - a 77% uplift on the prior year - with Microsoft alone committing $190 billion against the $152 billion analysts had previously mapped out.
Microsoft finance chief Amy Hood attributed $25 billion of that figure to higher memory and component costs rather than any grand new ambition, which is a measured way of saying the infrastructure stack is getting more expensive to hold at the same altitude.
Meta chief executive Mark Zuckerberg framed the acceleration differently when pressed by investors after the company lifted its spending range again.
"We are funding personal superintelligence to billions of people," Zuckerberg said.
Shareholders were unconvinced, with Meta's equity shedding ~9% on the day - the first genuine investor revolt against a capex trajectory that has only ever moved in one direction.
Jefferies analyst Brent Thill pushed back hard against the sceptics, telling the Financial Times the underlying economics remained sound.
"The bear thesis is garbage," Thill said.
Memory bottleneck
Here is the corner of the cycle that most coverage glosses over, which tends to be where the conviction sits.
The HBM manufacturers are arguably running a tighter book than NVIDIA itself, with SK Hynix posting a first-quarter operating margin of 72% - ahead of the GPU giant's own 65% - on revenue that nearly tripled year-on-year.
High-bandwidth memory and conventional DRAM are sold out across the board for 2026, contract pricing has ripped ~90% higher across a single quarter, and Goldman Sachs is calling it the sharpest supply crunch in 15 years.
The wry footnote is that the same squeeze inflating the memory oligopoly's profit spreads is adding ~14% to the retail price of your next handset - the consumer ends up steadily underwriting the compute boom, whether or not they own the stock.
The next frontier
The shape of AI compute is shifting beneath all this, and the rotation is moving away from the workload NVIDIA has cornered.
Inference - the execution of trained models, increasingly via autonomous agentic pipelines - now accounts for more than 70% of total accelerator cycles, and it happens to be the task that purpose-built bespoke silicon handles cheaply and at volume.
Custom accelerator shipments are forecast to grow 45% in 2026 against just 16% for merchant GPUs, with Google bifurcating its TPU architecture into discrete training and inference variants and Broadcom's AI revenues from that programme tracking toward $42bn by next year.
One projection has NVIDIA's inference allocation sliding from above 90% today to somewhere between 20% and 30% by 2028 - the kind of trajectory that should concentrate the mind of anyone pricing in permanence.
Huang closed the earnings call with a line that doubled as both a revenue declaration and a strategic thesis.
"Demand has gone parabolic. The reason is simple: Agentic AI has arrived," NVIDIA chief executive Jensen Huang said.
He had made similar ground at GTC earlier in the year, arguing that bespoke chips optimise for today's workloads while his GPUs are engineered for the one after that - a tidy defence that conveniently assumes the next architectural inflection always arrives on NVIDIA's timetable.
Bear call
The sceptics are not entirely barking at clouds.
Michael Burry holds put positions against NVIDIA and frames the whole edifice as a data-transmission bubble with the GPU giant cast as Cisco circa 1999 - pointing specifically at circular-financing arrangements, most visibly NVIDIA's pledge of up to $100 billion into OpenAI, a loss-making operation that promptly recycles capital back into NVIDIA silicon.
Bernstein Research analyst Stacy Rasgon put the structural concern plainly when the OpenAI deal landed.
"The action will clearly fuel 'circular' concerns," Rasgon said.
NVIDIA's counter, delivered via an internal seven-page rebuttal, is that such placements represent a fractional share of group revenue and the underlying draw is real - reasonable, until the moment private credit decides it no longer wants to be the next link in the chain.
What to watch:
- NVIDIA's 26 August print, and specifically whether guidance again outruns freshly revised Street models.
- Hyperscaler capex commentary through the second half, and any deterioration in free cash flow - Amazon's the one most likely to turn negative.
- DRAM and HBM contract pricing as the supply crunch extends toward 2027.
- The proprietary silicon ramp at Google, Amazon and Meta as a gauge of Nvidia's pricing power in inference workloads.
