Hyperscaler Debt, Token Prices Down 60%, and Why Berkshire's Meta and Google Buys Look Exposed

The Wall Street Journal is not panicking about whether Meta, Google, Microsoft, or Oracle can pay their bills tomorrow. Jose Luis Cava's cut is colder: solvency is not the issue. Viability is. The question is whether the returns on this wave of AI infrastructure will cover the interest and still produce an adequate profit.

The Debt You Do Not See Yet

Hyperscalers already carry large reported liabilities. Cava flags a larger pile off the balance sheet: data-center lease commitments that typically do not land as liabilities until the facility is operating. The headline debt understates the true obligation. That is not fraud. It is timing. Timing becomes a problem when the asset you are leasing has to earn its keep in a market where the unit price of intelligence is falling.

How the Token Machine Actually Bills

Language models do not sell "answers." They sell tokens.

A message is split into the smallest billable units of information. Input tokens are the user's prompt and context. Output tokens are the model's generated reply. Output is priced higher because it consumes more compute. Revenue is simply: volume of each type, times its price, summed.

That structure made the last two years look like a toll road. It also makes the business brutally sensitive to price per token.

Why Token Prices Fell About 60% in Two Weeks

Cava's alarm is not a vague "AI is over." It is a price crash in the unit the industry sells.

Customers got better at asking. After the first wave of oversized bills, firms learned to prompt leaner and consume less.

Compute got faster and cheaper. Efficiency plus competition compresses the price the vendor can charge.

China forced the floor. Open Chinese language models now deliver roughly 90% of frontier-model performance. That is close enough to trigger a global price war. You do not need to beat Gemini or ChatGPT on every benchmark. You need to be good enough, open, and cheaper.

When the unit price drops 60% this fast, the CapEx story stops being "how much capacity can we build" and becomes "will this capacity still pay for itself at the new price."

The Chart — and Berkshire's Average Cost

Cava's technical map is blunt.

Meta — rolled over after resistance near $700. The trend is now down. The measured target is the origin of the last up-leg, around April 2024, near $440 or lower.

Google — pulled back from $385. Base case: a slide toward $270, the March 2026 area.

Berkshire's problem is the average cost, not the brand.

• They bought Google in September 2025
• They added in Q2 2026 at an average near $350
• If the stock tags $270, the 2025 profit is gone and the 2026 adds are a loss

That is the kind of mark-to-market that pushes large holders into protective puts. Puts are not a vibe. They are selling pressure that can speed the decline they were meant to hedge.

November 3 Is the Aggravation Date

Cava's close: these names are damaged, and the situation can get worse from November 3. That date is already the political expiry on the Bessent–Japan liquidity architecture that has been pinning index calm into the midterms. A viability scare in hyperscaler returns does not need a recession to hurt. It needs a window where options expirations, forced hedging, and a 60% token-price reset are allowed to show up in the cash market.

The investors most exposed are those treating Meta and Google as "Buffett-approved, therefore done falling." Berkshire's cost basis is not a floor. It is a risk if the tape goes looking for $270 and $440.

---

This analysis is based on Jose Luis Cava's market commentary, August 18, 2026. For informational purposes only — not financial advice.