Nvidia projects 70% growth and signals the AI boom still has years to run

Nvidia released a rare forecast on Wednesday (27): it expects revenue to grow 70% in the next fiscal year, ending in January 2028. The number is significantly higher than the 44% projected, on average, by analysts ahead of the earnings report, according to Reuters via Economic Times. Shares rose nearly 5% in after-hours trading.

"AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable. Now, computing is revenue," said CEO Jensen Huang. It's an unusual disclosure: "We've never projected or given guidance a year ahead," he admitted.

What the numbers show

In the second fiscal quarter, ended in July, revenue more than doubled to $96.22 billion, above the $92.17 billion estimate. Adjusted profit was $2.22 per share, versus $2.10 expected. The data center division alone brought in $89 billion, also more than double the prior period.

For the third quarter, the company projects revenue of $108 billion (plus or minus 2%), above the $104.19 billion expected by consensus.

Nvidia's central argument is that the AI computing market is expanding, not peaking. "What makes the forecast even more credible is that demand is broadening beyond the original hyperscalers, with AI clouds, enterprises, sovereign buyers, and industrial customers growing materially faster," said Shay Boloor, chief strategist at Futurum Equities.

Vera Rubin and customer diversification

The new Vera Rubin platform, successor to the current generation, has already begun shipping to customers and is expected to account for about a fifth of data center revenue in the current quarter. The company also detailed where future demand will come from:

• AI labs (such as OpenAI) are expected to account for about a quarter of the business next year.
• The so-called neo-clouds, a category that includes Nebius and CoreWeave, are expected to end the year with more than 8 gigawatts of Nvidia GPU capacity, up from 3 GW at the end of last year.
• The partnership with Amazon Web Services was expanded: the two companies will deploy another 2 million Nvidia GPUs across Amazon's global infrastructure in 2027 and 2028.

The bottleneck holding back growth

Not everything is demand. Nvidia reiterated that it is supply-constrained. "Customer forecasts point to our growth doubling next year. However, we are constrained by supply," said CFO Colette Kress.

The critical point is memory. Rising component prices and higher costs are expected to pressure margins: according to Kress, gross margin should bottom out in the fourth quarter, between 71% and 72%, versus about 74% in the third. Analysts expected 74.77% in the third quarter.

On China, the picture remains murky. Washington cleared about 10 Chinese companies (Alibaba, Tencent, ByteDance) in May to buy the H200 chip, but shipments were stalled for months. Nvidia did not include China data center revenue in its forecast, a sign it is not counting on that market for now.

What changes for those building software in Brazil

For Brazilian developers, the practical takeaway isn't about the stock price, but about GPU availability and pricing. Nvidia itself admits it will remain supply-constrained, meaning demand for training and inference capacity will outstrip supply for quite some time. This has concrete effects on any stack that depends on GPUs:

• Expensive and contested instances. If neo-clouds and hyperscalers are fighting over capacity, anyone renting GPUs on demand on AWS, Azure, GCP, or providers like CoreWeave is likely to face queues, firm prices, and uneven regional availability. It's worth planning capacity reservations and considering architectures that don't rely on scaling GPUs elastically and immediately.
• Margin pressure turns into price. With Nvidia projecting that margins will fall because of memory costs, the bill eventually reaches the end cloud user at some point. Anyone budgeting an AI project for 2027 and 2028 would do well not to assume that the cost per token will drop linearly.
• Diversifying the inference strategy. The Vera Rubin cycle and the $89 billion in data center volume show that the market is structured around the CUDA ecosystem. Even so, for inference workloads (the part that most Brazilian teams actually run in production), evaluating alternatives such as managed APIs, smaller quantized models, and even competing chips can reduce exposure to this bottleneck.
• Sovereign and industrial buyers entering. Boloor cites "sovereign" buyers accelerating. It's a context in which national investments in AI infrastructure, including data center initiatives in Brazil, compete for the same global supply queue.

What remains open

The 70% forecast is a bet that the boom has "years to run," but it depends on two factors that Nvidia itself doesn't fully control: the memory supply chain's ability to keep up and the evolution of export rules for China. For tech teams in Brazil, the figure to watch isn't the guidance itself, but whether GPU scarcity will translate into higher costs and longer lead times throughout 2027, exactly the horizon in which many AI projects move from pilot to production.