Gavin Baker's 500K to 100M AI Leap Leaves Compute Starved for Air

Why 500,000 Users Can Still Mask a Compute Problem at 100 Million

500,000 users can still hide a broken math problem when the target is 100 million.

The core issue is not headcount or hype. It is what happens when an AI product moves from early adopters to mass usage. token cost efficiency is becoming the true competitive moat. At a small scale, a handful of expensive wins can look manageable. At scale, inference costs can overwhelm the model fast.

Cheaper AI can create a demand surprise

Bulls see a large market opening. Once AI becomes cheaper and good enough, usage can shift from niche tool to everyday utility. cheaper, smarter systems is the phrase already circulating in the market, and it captures the upside in one line: if intelligence per dollar keeps improving, usage can expand well beyond current patterns.

Where the bottleneck actually appears

The constraint is not interest. It is whether the stack can support mass-market scale without burning through economics on inference. At this point in the cycle, the limit is physical, not narrative. In practice, that means watts and wafers: power limits the footprint, and wafer capacity limits the silicon inside it.

That is why this matters now. Even if demand arrives, compute scarcity can throttle monetization before a business proves it can scale profitably. The harder investment question is not whether AI can attract users, but whether it can serve them cheaply and consistently.

TSMC, Power, and the Timeline That Matters

Demand can outrun supply once pricing falls

If AI shifts from expensive frontier models toward cheaper, smarter systems, the number of requests can rise quickly because the price per useful answer falls. But that revenue math only works if the hardware layer can handle the volume. If it cannot, the bottleneck stays in power infrastructure and fabs.

Why TSMC is the sharper pressure point

The wafer side is the tighter choke point because foundry leadership is hard to leapfrog. In that framework, the wafer bottleneck is TSMC is a useful shorthand for why supply may remain constrained even when demand becomes urgent.

Valuation debate and scarcity debate are related, but separate. The bullish case is that AI capex still looks defensible because the buildout is being funded largely by cash flows rather than fragile debt. The bearish concern is that the current phase still carries a temporary ROI divot, especially when massive training investments take time to pay back. The practical question for investors is whether supply can arrive fast enough to protect margins.

The near-term relief window changes the setup, not the present problem

Baker expects the near-term shortage to start easing in 2027 and 2028. That does not solve the current problem; it suggests the bind is contractual and infrastructural rather than permanent.

That creates a clear bull-bear split:

• Bulls can argue that the first builders who secure capacity will keep the upside once power comes online and supply widens.
• Bears can argue that relief arriving in 2027-28 is too late for companies that need margin proof now.

The key twist is economics, not storytelling. If AI adoption shifts toward cheaper models and market share moves away from the most expensive frontier inference, the ROI math can improve materially. Baker's own bull case depends on market share shifting away from certain frontier labs with 90%+ inference margins. Lower inference cost can support more users without requiring the same amount of scarce hardware per dollar.

What Investors Should Watch Next

The cleaner play is not to predict the final AI application winner. It is to track the binding inputs first. In this buildout, that means watching watts and wafers before the market settles on the downstream winner. If power infrastructure and advanced silicon supply stay tight, the scarcity trade can outlast the louder narratives.

Baker's practical angle is to follow TSMC's capacity decisions, because the wafer bottleneck is TSMC. From there, the next layer is which companies can allocate scarce compute efficiently rather than simply outbid others for whatever is available.

What would weaken the scarcity thesis

This setup changes in two straightforward ways:

• If the shortage starts to ease in 2027 and 2028 sooner than demand compounds, the scarcity premium should compress.
• If frontier labs keep 90%+ inference margins, the case for cheaper systems becomes less compelling because the economics of shifting away from them weaken.

The active disagreement is simple: does supply loosen first, or does cheaper demand arrive first?