Nvidia's 70% Guidance: AI Capex Is Far From Peaking

01 Trigger Event

Bloomberg reported on August 27 that Nvidia provided guidance of approximately 70% year-over-year revenue growth for the next fiscal year in its latest earnings communication, publicly hedging against the "AI capex peak" concerns that have accumulated in the market over recent months. Angelo Zino, Senior Vice President at CFRA Research, interpreted this on Bloomberg as blockbuster results.

The 70% figure itself is the news. Most companies build buffer into guidance—sandbagging is an unwritten rule of earnings season, especially when valuations are inflated by AI narratives. Nvidia's willingness to put out 70% directly shows their order book visibility extends well beyond a single quarter.

02 What This Really Means

The question isn't how many cards Nvidia sold, but why they dared to report this early and this high.

For the past two quarters, the market has been telling a story: after hyperscalers digest early Blackwell inventory, 2026 H2 capex growth will slow; inference workload ROI is unclear; enterprise AI budgets are tightening. This story has been told repeatedly, enough to push Nvidia's valuation down from its highs at one point.

The 70% guidance is Jensen Huang making a statement in capital markets language: we see no order cliff.

On a deeper level, this number is telling hyperscaler customers and model API providers that Blackwell/Rubin GPU supply tightness will persist for at least another 12-18 months. This directly affects two things: pricing power for model APIs, and the actual cost of self-built inference clusters.

I don't have first-hand information to break down this guidance—haven't seen hyperscaler order contracts—so I can't confirm how much of the 70% is already-signed LTA versus forecast. But Nvidia's CFO has consistently been measured in tone over the past several quarters; offering 70% proactively likely reflects locked-in long-term purchase agreements.

03 Historical Analogies

The closest parallel is the 1999-2001 fiber build-out cycle.

During the dotcom bubble, JDSU, Lucent, and Nortel burned through hundreds of billions of dollars laying transatlantic and transpacific submarine cables. When the bubble burst in 2000-2001, most demand-side companies died—but the fiber remained on the ocean floor. From 2003-2007, when new demand emerged from YouTube, Netflix, and cross-border enterprise SaaS, that surplus fiber suddenly became a scarce asset, not because new cable was laid, but because it had been laid years earlier.

AI infra is now in a similar position: the capacity Huang is building may not all correspond to current end-demand. But if agent workloads, enterprise internal inference, and robotics inference scenarios actually materialize in 2027-2028, pre-locked compute will be a structural advantage.

Another parallel is AWS pre-2014. Amazon ran capex growth above 40% for over a dozen consecutive quarters, with capital markets berating Bezos for burning cash for years—until AWS profits were released and the stock was repriced. Nvidia today enjoys treatment AWS never did: capital markets are at least willing to believe AI capex will deliver returns.

04 What This Means for AI Builders

Decisions to adjust this week:

First, expectations for the inference cost curve. If Blackwell/Rubin supply tightness extends into mid-2027, the commonly-cited assumption that "inference prices will drop another 50% over the next 12 months" may be overly optimistic. Model API wholesale prices, batch discounts, and prompt caching incentives will all be squeezed tighter. Unit economics designed around "inference cost approaching zero" need to be re-run.

Second, the value of compute lock-in is rising. Players holding long-term GPU supply agreements (CoreWeave, Lambda, Neocloud types) are seeing their moats widen. Several AI infra middle-layer players I've engaged with previously arbitraged spot capacity; now they're more focused on whether they have 18-24 month fixed capacity commitments. This is a structural shift, not a cyclical fluctuation.

Third, the logic for model API selection needs recalibration. Anthropic, OpenAI, and Google all have large-scale GPU reserves themselves; their API price stability will exceed that of mid-tier players dependent on external compute. For token gateways like opc.club, part of the value is hedging compute supply risk—this layer of logic needs to be articulated clearly in product.

Fourth, early signals on the training side. If Nvidia's 70% is real, next-generation frontier model training clusters will continue scaling up. This means pretraining's capital threshold rises further, and the window for open-source models to catch up to frontier is narrowing.

The above judgments rest on the premise that the 70% guidance is roughly accurate. If subsequent earnings reports show this was overly optimistic, the entire logic needs to be rewritten.

05 Counterarguments / Risks

I may be overly optimistic in several places. Let me be explicit.

First, concentration risk in the 70% growth. Nvidia's revenue structure is heavily concentrated in the top 4 hyperscalers. If one or two significantly cut external procurement in 2027 due to accelerated progress on their own TPU/Trainium/Maia chips, this 70% collapses immediately. Microsoft Maia, Google TPU, AWS Trainium are not paper projects—I haven't seen their production timelines internally, but the direction is clear.

Second, sovereign AI exposure. A significant portion of Nvidia's growth over the past two years has come from Middle East sovereign funds (UAE G42, Saudi HUMAIN) and Southeast Asian national projects. This demand is extremely sensitive to geopolitics—once export controls tighten again (US State Department/BIS have moved several times in 2025), this revenue evaporates instantly.

Third, Nvidia itself may be repeating Cisco's 2000 mistake. Cisco at the time also operated under the narrative that global network construction would never stop—when in fact ISP first-wave build-out was peaking. I'm not saying Nvidia is Cisco, but when forward guidance is set too full, historical cases of management errors are not rare.

Fourth, the worst case: the 70% growth is real, but gross margin slides from 75% to below 60%. This kind of revenue-for-margin expansion gets double-punished in an environment of multiple compression. I didn't see this broken down in this report, but the next earnings will need to focus on gross margin trajectory.

If any of these four risks materialize, all the judgments above need to be discounted. But even so, the 70% figure remains an industry-level statement: Huang believes AI infra build-out is far from peaking. This judgment itself, more than any single-quarter revenue number, deserves AI builders' attention.