NVIDIA Could 10X Its Sales. One Bottleneck Is Stopping It

Gavin Baker, the tech investor whose track record gives him an audience whenever he sits down with Patrick O’Shaughnessy, floated a number on a recent podcast that should make anyone modeling NVIDIA’s (NASDAQ:NVDA | NVDA Price Prediction) future do a double-take. NVIDIA’s chip sales, he argued, could grow to $2 trillion, possibly $2.5 trillion, possibly $3 trillion by 2027. The company sold roughly $215 billion in chips last year. The high end of Baker’s scenario is roughly a 10x increase.

The interesting part is what Baker thinks is preventing it. A single constraint, in his view, stands between NVIDIA and that outcome: wafer supply from Taiwan Semiconductor. And Baker says he is glad about it. If NVIDIA could build as much as it wanted, AI capex would tip into bubble territory. The bottleneck is the circuit breaker.

What NVIDIA’s order book already looks like

NVIDIA posted yet another beat in Q1, with EPS at $1.87 adjusted vs. $1.76 estimated. Revenue came in at $81.62 billion vs. $78.86 billione stimated. These figures exclude any China data center compute. Data Center revenue nearly doubled, with hyperscalers accounting for half of it.

The demand pipeline behind Baker’s math is concrete. Meta has signed a multiyear deal for millions of Blackwell and Rubin GPUs. OpenAI has contracted for 10 gigawatts of NVIDIA systems. Anthropic has committed to 1 gigawatt, CoreWeave 5 gigawatts by 2030. Jensen Huang’s read on Q3 was succinct. “Blackwell sales are off the charts, and cloud GPUs are sold out.”

You can see the customer pull in the balance sheet too. NVIDIA carried $95.2 billion in supply-related commitments at the end of Q4, money already promised to vendors to lock in wafers, packaging, and HBM. Wall Street’s analyst target sits at $275.31 against a stock that has risen 62.77% over the past year. The forward multiple, 27x, looks almost restrained given a profit margin of 55.6%.

Why Taiwan Semiconductor is the gating factor

That demand has to land on a wafer somewhere, and there is essentially one address. Taiwan Semiconductor (NYSE:TSM) holds roughly 72% of the leading-edge foundry market. On the Q1 call, CEO C.C. Wei made the timing problem explicit. “It takes 2 to 3 years to build a new fab,” he said, “and it takes time to ramp it up.”

TSMC is responding with capital. The board approved roughly $31.28 billion for advanced technology capacity and fab construction, plus up to $20 billion injected into TSMC Arizona. 2026 capex is now guided toward the high end of $52 billion to $56 billion. Wei’s specifics on N3, the node that bakes most of Blackwell and Rubin, tell you when relief arrives. Tainan’s new N3 fab hits volume in the first half of 2027. Arizona’s second fab ramps in the second half of 2027. Japan’s second fab follows in 2028.

The other constraint sits in advanced packaging. CoWoS, the technology that fuses GPU dies to high-bandwidth memory stacks, has been the industry’s tightest pinch point for two years. Wei said TSMC is now supplying the largest reticle size CoWoS and has built a CoPoS pilot line for production “a couple of years later.”

Insiders are putting their own money behind that story. On May 8, 25 TSMC executives, including Chairman and CEO C.C. Wei and both Co-COOs, bought common shares at $71.82. A nearly identical group bought on April 9 at $57.87. Zero common-share selling in the period. TSM is up 105.34% over the past year.

The bottleneck as circuit breaker

If NVIDIA could triple shipments tomorrow, hyperscalers would absorb it, build the data centers, and finance the whole thing with leverage. That is the recipe for a 1999-style overshoot. Wei describes the AI accelerator growth rate as “toward higher 50s of CAGR” through 2029, which is enormous but paced. TSMC’s two-to-three-year fab cycle forces hyperscalers to commit capital against actual silicon delivery windows, not vibes.

For investors, the read-through is straightforward. NVIDIA’s revenue ceiling in any given year is whatever TSMC can ship. TSM is the throttle. Watch the CoWoS expansion cadence and the N3 ramp dates, because those are the inputs to Baker’s $2 trillion to $3 trillion math.