Silicon Valley is spending billions on AI tokens and nobody can agree if it's working

The article highlights growing skepticism around aggressive AI token usage, with Uber COO Andrew Macdonald saying he has not seen direct productivity gains from higher token consumption. Google CEO Sundar Pichai warned the problem of companies blowing through AI budgets may worsen, while reports cited Uber exhausting its annual AI budget in just the first four months of the year. The piece suggests rising concerns about wasted spend, weak ROI, and potential AI bubble risk, though it is more of a sentiment/efficiency debate than a direct market-moving event.

Analysis

The first-order read is not “AI is broken,” but that the market is transitioning from an enthusiasm regime to a productivity-accountability regime. That matters because the near-term spend curve for model providers can decelerate even if unit adoption keeps rising: enterprises will keep using AI, but they will shift from open-ended experimentation to metered, workflow-tied deployment. The result is likely a compression in low-ROI inference demand and a relative reallocation toward vendors that can prove software-level productivity gains rather than raw usage growth.

The biggest second-order loser is Nvidia if the market starts marking down token growth as a proxy for monetizable demand. GPU utilization remains strong today, but the narrative risk is that CFO scrutiny forces customers to optimize prompts, cache outputs, route less work to frontier models, or move marginal workloads to cheaper alternatives. Over a 3-6 month horizon, that can slow incremental spend growth even while headline AI adoption stays high, which is exactly the kind of change that multiple-sensitive semis tend to discount first.

Google is the cleanest “quality over hype” setup because it can benefit if enterprises start demanding measurable productivity and cost discipline. A budget-conscious enterprise buyer tends to favor bundled, managed, and cheaper-to-deploy stacks over bespoke token-intensive workflows, which supports GOOGL’s distribution advantage. Visa is a quieter beneficiary: AI governance pushes firms to instrument workflows more tightly, and the companies that monetize AI most effectively will likely be those with robust transaction-level measurement and enterprise workflow control, not the loudest token consumers.

The contrarian point is that the backlash may be more about accounting than economics. If a lot of internal token spend is currently waste, cutting it can actually improve margins and extend the runway for useful AI adoption rather than kill demand. That argues for fading the most crowded “AI usage = growth” trade and favoring businesses with pricing power, embedded distribution, or direct ROI proof. The main catalyst that would reverse the short case on NVDA is evidence that enterprise optimization is expanding total addressable workloads faster than it reduces token waste, which should show up first in cloud commentary and capex guideposts over the next 1-2 quarters.

Key Decisions for Investors

• Short NVDA on rallies over the next 2-6 weeks; use a call spread or 1-2 month put spread to express narrative compression risk. Risk/reward favors downside if enterprise AI spend normalizes before the market reprices capex growth expectations.
• Long GOOGL vs short NVDA in a pair trade for a 3-6 month horizon. The thesis is that cost-conscious AI adoption favors integrated distribution and lower-cost inference rather than pure hardware leverage.
• Add to V on pullbacks over 1-3 months. If enterprises tighten AI governance, the winners are the platforms that sit inside workflow and payments rails, where measurement and control are easier to monetize.
• Underweight UBER and META tactically for 1-2 quarters if you expect internal AI budget scrutiny to hit operating expense narratives before it lifts top-line productivity. Any long should be sized as a quality-growth hold, not an AI-beta trade.
• Watch GOOGL and JPM commentary into the next earnings cycle for language around AI cost discipline versus output gains; if management teams start tying AI to concrete throughput metrics, rotate into the names with verifiable ROI and away from token-growth proxies.