Cognition CEO says tech companies got ‘carried away’ with token leaderboards and should measure employees on output instead

Cognition CEO Scott Wu says “tokenmaxxing” has gone too far: some firms are ranking engineers by token usage rather than output, and internal token leaderboards at Meta/Amazon were reportedly scrapped after bots completed useless tasks. The article cites Uber burning its entire 2026 AI budget in four months and Amazon capping token spending at $1,500/month, despite token prices dropping ~90% since 2023—highlighting a lopsided spend-to-output problem. BCG data also shows AI saves time (8 hours/week reported by 42% of workers) but 66% received little guidance and half didn’t redeploy saved time to strategic projects, implying cautious, ROI-focused AI deployments are needed.

Analysis

The signal here is not “AI adoption is failing”; it’s that procurement is shifting from usage theater to measured productivity, which is bad for any business model or stock narrative built on rising token consumption without a hard ROI loop. In the near term, that creates a sentiment headwind for high-multiple AI-adjacent names and for cloud/infra vendors that still rely on expanding consumption rather than clear labor substitution. META and AMZN are most exposed on the narrative side because investors are already sensitive to capex intensity and internal AI sprawl; if management teams start talking more about quotas, gates, and ROI thresholds, the market will read that as evidence that marginal AI spend is being cut first.

The second-order effect is more nuanced: low-value token traffic gets culled quickly, but the workloads that survive should be stickier and more valuable. That favors enterprises with disciplined deployment and measurable throughput gains—GS is a mild beneficiary if it can demonstrate real engineering/ops leverage—while leaving “AI as a workplace perk” models vulnerable. UBER is a caution flag: budget caps are a negative only if AI is still being treated as an uncontrolled opex line; otherwise they are margin-positive and should reduce the chance of another surprise spend overrun.

Catalyst path is 1-3 months: next earnings calls, budget resets, and any disclosure of AI productivity metrics will determine whether this is a one-quarter cleanup or a broader spending discipline regime. Over 6-18 months, the market should reward vendors and users that can tie AI to unit economics, not engagement. The contrarian read is that the consensus may be over-discounting AI demand: trimming waste does not equal abandoning the category, and the real beneficiary is the subset of companies that can prove output per dollar.

Key Decisions for Investors

• Short META on rallies over the next 1-3 months; thesis is multiple compression if AI remains framed as internal experimentation rather than monetized output. Cover if management quantifies clear ad/engagement uplift from AI tools or if capex-to-revenue efficiency improves materially.
• Short AMZN tactically into earnings/capex commentary if cloud AI consumption growth decelerates while spending stays elevated; the risk is that AWS keeps winning the highest-value enterprise workloads even as low-value token traffic is cut.
• No standalone trade on UBER yet; treat the spending cap as a watch item. If the next update shows AI budget discipline without service degradation, the market should move from ‘overspend’ penalty to margin-credit, which would falsify a bearish read.
• Mild relative long GS vs. META over 3-6 months as a ‘measured productivity’ versus ‘AI optics’ pair; only size this if GS can show AI-driven operating leverage in results, otherwise the edge stays too small.