Nvidia's 70% Growth Bet Hides Its Deepest Risk

The number that isn’t about confidence

Nvidia guided for 70 percent revenue growth in fiscal 2028. It blew past S&P Global’s 44 percent consensus estimate and came on the heels of a second quarter that delivered $96.2 billion in revenue — up 106 percent year over year, beating every street guess.

On the surface, this is the kind of print that shuts down bubble narratives for a quarter. Jensen Huang called it cleanly: AI has reached an inflection point, and demand is “near 100 percent.” The 70 percent forecast, he said, is not a ceiling on what the market wants — it is a ceiling on what Nvidia can physically deliver.

Read that sentence the other way and the real story appears.

Nvidia is not just telling the market it can grow fast. It is telling the market it has run out of room to grow faster. The constraint is no longer demand. It is supply — silicon, memory, packaging capacity, and the alliances that make all of that real.

Why the supply constraint is the interesting constraint

The quarterly results contain a detail that most summaries will flatten into a revenue headline. Nvidia disclosed that its purchase commitments with suppliers have more than doubled in three months — the largest increase in company history — and for the first time explicitly noted that most of those commitments are tied to memory and manufacturing facilities.

Collette Cruse, the CFO, said the current memory shortage is “largely driven by AI infrastructure expansion.” Nvidia is working closely with its three major memory suppliers: Samsung, SK Hynix, and Micron. On the day of the earnings release, Samsung and SK Hynix shares rose 1.72 percent and 2.49 percent respectively.

This is a company racing to buy the inputs it needs before someone else buys them first. That is not the behavior of a firm whose only problem is convincing customers to spend. That is the behavior of a firm whose customers are spending so aggressively that the supply chain itself has become the bottleneck — and the competitive front line.

The data center math is brutal and seductive

Here is the economic engine behind the guide: a data center equipped with Nvidia’s Hopper (H100) chips, launched in 2022, generated roughly $18 billion in revenue per gigawatt of capacity. The same metric for Blackwell, released in 2024, sits at $250 billion per gigawatt. Nvidia expects the Vera Rubin generation, arriving in 2026, to reach $400 billion per gigawatt.

That is a tenfold increase in four years. It is also the reason every hyperscaler on earth is trying to sign an Nvidia deal before the quarter ends.

The per-token economics work in Nvidia’s favor. Cruse noted that Vera Rubin delivers 30 times the processing throughput of the prior generation while cutting token cost to one-thirty-fifth. Huang said the tokens generated by customer GPU clusters produce “unbelievably high returns.”

But the math cuts in two directions. If a single data center can generate $400 billion per gigawatt, then the return on infrastructure investment is enormous — and the incentive for every cloud provider to lock in Nvidia supply is existential. The flip side is that the capital required to build that infrastructure is also enormous, and the payback period is compressing faster than the balance sheets of some buyers can absorb.

The AWS deal is a signal, not a distraction

Nvidia announced an additional supply of two million GPUs to Amazon Web Services over two years. The deal is large by any conventional measure. It is also a quiet admission of something important: Nvidia cannot grow at 70 percent without AWS buying far more than it planned a year ago.

That is not a problem today. It is a problem in 2029, when the two-year commitment expires and AWS — along with Google, Microsoft, and Meta — has had time to ship its own silicon into production and prove it can handle a meaningful slice of inference workloads.

Huang pushed back hard on the de-Nvidia-fication narrative. He pointed out that many custom chips, including OpenAI’s recently announced Halipeño, are inference-optimized accelerators designed for a single cloud environment, whereas Nvidia’s GPUs remain general-purpose across every cloud. The distinction is real. Custom chips are cheaper at the margin for their intended workload. They are also narrower. The question is whether narrow wins over time as models stabilize and workloads fragment.

Sovereign AI and the customer base that complicates the picture

Perhaps the most underreported part of the quarter is the AICE segment — AI Cloud, Sovereign AI, and Enterprise — which grew 138 percent year over year, outpacing the overall 106 percent increase. Huang said roughly half of Nvidia’s market now sits outside the hyperscalers, in neocloud operators and sovereign AI programs.

Sovereign AI is the strategic wildcard. Governments in Europe, the Middle East, and parts of Asia are building nationally controlled AI infrastructure, often with Nvidia hardware at the core. That creates a new class of buyer with different incentives than a hyperscaler: less pressure to vertically integrate chips, more pressure to ensure data residency and regulatory compliance, and a funding model backed by state budgets rather than quarterly margins.

It also means Nvidia’s customer base is diversifying away from the very companies most likely to develop in-house alternatives. A hyperscaler builds its own chip to reduce dependency on Nvidia. A sovereign buyer has less reason to do so — and more reason to deepen it.

The investment bet that looks like strategy

Nvidia disclosed it has invested roughly $50 billion in companies like Anthropic and OpenAI. Huang called the amount “a tiny fraction” of the company’s free cash flow and said he regrets not investing more, earlier. He framed the investments as support for a virtuous cycle in which Nvidia supplies the infrastructure and the applications companies supply the demand.

That framing is carefully constructed. The virtuous cycle is real — more AI applications drive more GPU demand, which drives more Nvidia revenue, which funds more R&D and more supplier commitments. But virtuous cycles have a blind spot: they assume the cycle participants remain dependent on the cycle’s central node.

Anthropic and OpenAI are dependent on Nvidia today. They are also, by definition, companies whose entire business exists to make advanced AI more capable. As those capabilities improve, the economic case for custom silicon strengthens. The $50 billion in investments is a hedge against that outcome — a way of retaining influence over the companies most likely to try to leave.

It is also a sign that Nvidia’s leadership understands the dependency runs in both directions.

What the 70 percent guide really means

A 70 percent growth rate is extraordinary. It is also a commitment to keep expanding supply faster than the industry has ever seen, while managing relationships with customers who are simultaneously growing more powerful and more capable of building alternatives.

The memory shortage is the most visible symptom of that tension. The purchase-commitment doubling is the company’s response. The AWS deal is the proof that demand still vastly exceeds what can be delivered. The AICE growth rate is evidence that the customer base is widening even as the core hyperscaler relationship grows more complicated.

Nvidia is winning. The question this quarter raises is not whether it will keep winning — the guidance says it will — but what happens to a company whose dominance depends on keeping its customers hungry while quietly training them to be less dependent on it.

The 70 percent number is not a bubble deflator. It is a countdown clock.