Five Charts That Explain Nvidia's Ten-Day Win Streak

The data centre revenue line tells the entire growth story. From $10 billion in 2021 to $155 billion in 2025 is the fastest scaling of a hardware business in the history of publicly traded semiconductors. The 2026 consensus sits at roughly $200 billion, implying another 30% year-on-year expansion.

The deceleration is real and it matters, but the absolute dollar growth is still the largest in the industry. The $45 billion of incremental revenue that 30% growth represents is larger than AMD's entire data centre revenue base, and larger than the entire revenue of most standalone chip companies. The scale of what Nvidia is doing at this growth rate has no precedent in the sector.

Historically, hardware businesses that cross the $100 billion mark in a single segment have seen the growth rate converge to 15-20% within three years. That is the natural law of large numbers. Nvidia is fighting that gravity with a product cycle that continues to pull the curve forward.

The gross margin is the quiet miracle. Hardware businesses at scale do not typically hold 74% gross margins. The closest historical parallel is Intel in the late 1990s, and that margin structure proved cyclical. The current Nvidia margin reflects pricing power on the Hopper and Blackwell generations plus the memory cost deflation that has helped the supply chain.

A compression back to 70% would be a meaningful earnings hit. A compression back to 65% would be existential for the current multiple. The question the market has to answer is whether the margin structure is now secular, driven by irreplaceable CUDA software lock-in, or cyclical, driven by a temporary supply-demand imbalance that custom chip competition will eventually close.

The Blackwell Ultra roadmap suggests Nvidia can hold the line on margins into 2026. The 2027 generation is where the pressure points start to show up, and the custom chips from the hyperscalers are the most immediate threat to the margin line.

Eighty-two billion of free cash flow at a 35% conversion rate from revenue is the cleanest fundamental data point in the entire AI trade. The capital return programme that sits on top of this cash flow has retired more than 4% of the share count in 2025 alone. At current prices the buyback is still accretive, even at the elevated multiple.

The forward question is whether the cash flow supports the multiple. At $3 trillion market cap, the current FCF yield is roughly 2.7%. That is defensible if growth continues at the guided pace. It is vulnerable if revenue growth falls below 20% while capex for the next-generation fab commitments remains elevated.

Nvidia's capex is still a fraction of the hyperscaler capex, because Nvidia outsources manufacturing to TSMC. That gives Nvidia a structurally higher free cash flow conversion than the hyperscalers who are building the data centres that use Nvidia's chips. The cash flow dynamic is the reason the buyback programme has been able to compound at the pace it has.

The quantum bet that Nvidia flagged this week with the latest AI move is the newest growth vector. The TAM is small today and the monetisation is early, but the infrastructure investment is real. Historically, when a dominant compute platform extends into an adjacent TAM with the capex headroom Nvidia has, the adjacency captures a share premium over its competitors.

This is not a 2026 number. It is a 2028 and 2029 story. But it is a genuine incremental growth vector, and the market is starting to price a modest contribution. Robotics, automotive compute, and edge inference are the other adjacencies that the CUDA software stack extends into cleanly.

The AMD comparison is instructive. AMD has a credible GPU roadmap but does not have the software ecosystem to capture the adjacent TAMs. That is the part of the moat that the market still underestimates in Nvidia.

This is the chart that has us trimming. The forward sales multiple has re-expanded into the zone where every previous rally has topped. Each of the last three 10-day win streaks has been followed within six weeks by a drawdown of at least 12%. The base rate is unfavourable.

The multiple is not a prediction tool on its own. But the combination of rising multiple, decelerating growth rate, and reaccelerating competitive pressure from custom AI chips at the hyperscalers is the pattern that has preceded every prior pause. The data does not require the bull thesis to be wrong. It just says the entry point here is poor.

Historically, when a dominant technology stock trades at the upper band of its forward sales multiple during a decelerating revenue cycle, forward six-month returns have been negative in seven of the last ten observations across the large-cap semiconductor peer set.