Three Chip Cycles Point to the Same Conclusion on Nvidia's Physical AI Push

Nvidia's numbers are extraordinary. Revenue compounded from $26.9 billion in FY22 to $215.9 billion in FY26. That is a 700 percent increase in four fiscal years. Net income moved from $9.8 billion to $120.1 billion, a 1,225 percent increase. Free cash flow followed a similar curve, from $8.1 billion to $96.7 billion. The efficiency of conversion, FCF of $96.7 billion on revenue of $215.9 billion, implies a 44.8 percent FCF margin. Intel's best year in its entire operating history produced an FCF margin below 20 percent. This is not a normal operating company.

The Research Desk's framing is simple. Nvidia has caught the first true hyperscaler capex cycle in computing history. Training cluster demand has been the single largest buying event in enterprise hardware since mainframes. The scale of the revenue reset, and the speed at which it happened, places the current period alongside 1996, 2000, and 2018 on the shortlist of true semiconductor inflections.

At the same time, the operating margin of 65 percent is higher than any of those prior cycles peaked at. The 1996 cycle peaked with Intel at roughly 35 percent operating margin. The 2000 cycle peaked with JDSU at 38 percent before collapsing. The 2018 crypto cycle peaked with Nvidia at 40 percent. A 65 percent operating margin is a very specific signal about pricing power. It is also a very specific signal about mean reversion risk when the cycle turns.

What distinguishes Nvidia today from each of those prior cycle peaks is the software moat. CUDA is not easily replicated and the developer base using it has grown to millions. That is a real defensive asset. The question is whether the moat is wide enough to prevent the margin reset that historically accompanies this combination of revenue growth and customer concentration.

The 1996 Intel cycle is instructive because it produced a two-leg revenue expansion rather than a single peak. The first leg was the 1996-1997 Pentium II ramp, which generated $12 billion of revenue uplift and drove the stock from $14 adjusted to $41 adjusted. The second leg was the 1998-1999 enterprise server cycle, which added a further $15 billion of revenue and drove the stock to $85.

The reason the second leg materialised was customer base broadening. Intel's top four customers dropped from representing 52 percent of segment revenue in 1996 to 34 percent in 1998 as enterprise IT took on the buying load. That broadening supported the second capex wave, which is what the second equity leg was paid for.

For Nvidia, the analogue to that broadening is the shift from hyperscaler training workloads to enterprise inference workloads and in the end to what the CEO has labelled Physical AI. If the next two years show enterprise inference revenue reach 30 percent of datacentre total and Physical AI produce at least a first meaningful segment disclosure, the cycle extends. If the concentration stays where it is, the analogy to Intel 1996 breaks and the stock is priced for a cycle that does not have a second leg.

The Intel 1996 parallel also produced a specific lesson about how multiple compression begins even when revenue is still growing. The 1999 consensus for Intel's FY00 earnings was cut twice inside the year, and the multiple compressed two turns before the revenue line even rolled over. Multiple compression leads revenue. Investors waiting for the revenue miss to sell are typically selling into a stock that has already re-rated. That sequencing is the single most important lesson the Research Desk draws from the 1996 cycle.

The 2000 optical cycle is the cautionary tale. JDSU and Ciena produced revenue doublings inside eighteen months. Both stocks re-rated to multi-hundred forward multiples. The cycle was real in the sense that fibre infrastructure was genuinely being built. The problem was that operating margin at the peak was impossible to defend because customer inventory builds were larger than true end demand.

When inventory draws began in early 2001, revenue halved in two quarters. Operating margin compressed from 38 percent to negative 15 percent because the fixed cost base could not flex down fast enough. The equity collapsed 85 percent in the following year. The lesson the desk takes from the 2000 cycle is specific: the risk at a peak is not revenue reversal, it is operating margin reversal. A 600 basis point margin compression on a doubled revenue base moves earnings down 50 percent. That magnitude of compression happened inside six months.

For Nvidia, the margin risk sits around the training-to-inference mix shift. Training hardware carries materially higher gross margin than inference hardware. If the shift to Physical AI or broad enterprise inference accelerates faster than expected, revenue may grow but margin may compress meaningfully. That is a different risk to a pure cyclical reversal but it shows up the same way in the earnings line.

The additional warning from 2000 is that the optical cycle's end was not announced by a single customer cancellation. It was announced by three separate data points arriving within ninety days of each other: a telecom capex guide down, a component supplier inventory warning, and a major customer order push. When those three data points arrive in sequence for Nvidia, the cycle is turning. We are watching for all three.

The 2018 Nvidia crypto cycle is the most recent parallel, and the one many bulls cite when arguing the current cycle does not end like the 2000 one did. In 2018, crypto mining demand added roughly $1.5 billion of quarterly revenue. When that demand withdrew in late 2018, quarterly revenue dropped roughly $900 million and gross margin compressed 400 basis points. The stock fell 52 percent peak to trough across six months.

What made the 2018 cycle recoverable was that gaming and datacentre demand continued to grow through the crypto withdrawal. Customer base concentration was 45 percent at the peak but broadened modestly as enterprise deep learning workloads started ramping in 2019-2020. By mid-2020, the stock had re-rated back above the 2018 peak.

The lesson the desk takes from 2018 is that Nvidia's business has demonstrated a tolerance for pullbacks when the underlying compute demand continues to broaden. The current cycle will have its withdrawal event, whether from hyperscaler digestion or from custom silicon substitution. The question is whether Physical AI and inference workloads are ready to absorb the pullback when it comes. The 2018 reference point is encouraging but not determinative. Markets rarely trade the same pattern twice without adjusting for the scale difference. A 52 percent decline from today's market cap is $2.6 trillion of equity value, which is a different magnitude of pain than the 2018 reset even at the same percentage change.

Nvidia's current valuation is generous on every normal metric. Price to sales of 22.8 is roughly twice the semiconductor cohort average. Forward PE of 24.5 is reasonable if earnings continue to double, but consensus models a significant deceleration. Trailing PE of 41 is inflated relative to any realistic mid-cycle earnings base.

The structural question is what mid-cycle earnings look like. Using the FY26 earnings base of $120 billion as a fully-priced cycle peak, and applying a modest 15 percent margin compression scenario combined with revenue growth of 20 percent CAGR through FY28, the resulting FY28 earnings base is roughly $155 billion. At 20x that number, fair value is $3.1 trillion market cap against the current $4.92 trillion. That is a 37 percent gap.

The bull case requires the margin compression to not happen and revenue growth to exceed 20 percent CAGR. The Research Desk's base case is that the margin compression is mechanical given the mix shift to inference and the entry of custom silicon from hyperscalers. We model FY28 earnings closer to $140 billion and fair value closer to $2.8 trillion. The current price embeds the Physical AI upside as already delivered.

Relative to the 1996 Intel peak, the 2000 JDSU peak, and the 2018 Nvidia peak, the current multiple is below the peak multiple of each of those prior cycles. That is the bull's correctly-argued point. What it misses is that the absolute margin starting point is much higher this time, which means the mean reversion path is steeper even if the multiple starting point is lower. The math favours caution here.

The single biggest threat to Nvidia's revenue concentration is custom silicon development at the hyperscalers. Google has TPU generations that are maturing on a two-year cadence. Amazon has Trainium and Inferentia lines that are increasingly competitive for specific workloads. Meta has disclosed custom ASIC development for inference. Microsoft has Maia silicon on a roadmap.

Each of these programs takes years to scale and none has displaced Nvidia silicon at the leading edge. But at the margin, they represent a customer strategy that looks nothing like the strategy Intel's customers pursued in the 1996 cycle. Customers who are also developing their own silicon have a structurally different interest in the pricing power of their vendor. The 65 percent operating margin is a direct invitation for customers to accelerate custom silicon programs.

The Research Desk's view is that custom silicon will absorb roughly 20-25 percent of the hyperscaler training workload by FY29. That is not catastrophic for Nvidia revenue, but it does cap the pricing power and compresses the incremental margin on new shipments. The margin peak has more informational content than the revenue trajectory.