Motivation: Hyperscalers are locked in a speed race for intelligence 

As demand has outstripped supply, and unit economics of tokens have been trending to 0, induced demand causes GPU shortages. GPUs are, as a whole, extremely inelastic in supply, which means significant spend on GPUs cause hyperscaler capex to grow 2-3x y/y. 

So this shortfall is obviously a problem for the hyperscalers. As demand for intelligence exponentially increases, and GPU costs skyrocket, this rebound effect shows up more pronounced on hyperscaler cash flow statements.

We can see then that the capex squeeze is starting to hit balance sheets of the hyperscalers. 

Note though that regardless of the runaway capex spending, let’s take a look at balance sheets here, short term liquidity remains pretty strong.

Q2 2026, $B	Alphabet	Meta	Microsoft	Amazon
Cash + near-term securities	242.5	90.3	76.8	123.0
Cash-like, ex. marketable equities	155.4	86.7	76.8	~118.3
Current liabilities	126.1	56.4	168.8	241.3
Q2 capex / infrastructure cash spend	44.9	31.1	35.8	53.1
Q2 free cash flow	–5.9	+0.8	+19.6	~–7.7
Cash-like ÷ Q2 capex	3.5 quarters	2.8 quarters	2.1 quarters	2.2 quarters

The final row is a very aggressive estimate of cash burn assuming that operating cash flows go to zero. The scary thing about this cash burn is not yet “these companies will run out of cash” but instead “at the current rate of consumption, these companies with Billions of liquid resources can only self fund for a few quarters with current capex if their operating cash flow stops covering the bill.” 

There are two assumptions baked into this model.
1. FCF bleed in Capex for the hyperscalers remains constant. There’s no evidence so far to validate this. As we’ve shown above, as token costs go down, compute demand goes up, and costs have increased by multiples over that exact period.

2. The physical chokepoint in Taiwan. Geopolitical tail risks are always a problem, and are very hard to price.

This is a pure induced demand problem. The better AI gets at intelligence, the more it’s used, the more spending is required to access intelligence, the more each hyperscaler needs to eat at its operating budgets to serve AI.

Markets for Intelligence: Socializing Compute in a Supply-Constrained World

Compute markets transform concentrated risks rather than simply eliminating them: they can diffuse ownership, financing and access risk while simultaneously increasing utilization and propagating strategic race dynamics across the system.

This statement motivates a mechanism design question: Can we socialize compute without socializing the race?