Alok Agrawal

I'm not privy to what she's seeing, but if I had to guess: (a) she has to sound bullish, and (b) spot is running at 3–5x contracted pricing, so it's possible that is what she was alluding to as opposed to realistic long-term contracts. Can't capitalize those at more than 3-4x free cash flows given how supernormal profits are at that pricing. This isn't fundamentally different from how any commodity supercycle plays out. What I find most interesting about this point in the cycle is that the vast majority of token consumption has come from coding. Most established software and enterprise companies are using it to work down tech debt and accelerate their roadmaps. Yet software equities are trading as if these businesses are close to ex-growth. Something has to give. AI adoption in coding has followed an unusually steep S-curve because the ROI is enormous and the technology works. Developers are also the most sophisticated users, and they can quickly catch and correct agent errors. But what if that S-curve plateaus early precisely because adoption was so steep? Are people even thinking about that? Anecdotally, I'm also seeing development processes being re-architected to use expensive models for planning and cheaper models for implementation. That lowers revenue per unit of coding work even as volume grows. If 70% of revenue stops growing rapidly, the remaining 30% has to grow roughly 2.5x faster to compensate. And we know there are real impediments to adoption of both consumer agents (Muse growth is already slowing) and enterprise agents (Satya covers some of these in his recent memo). Ultimately, token revenue has to come through to support current capex projections of over $1T next year. That requires $1.5T+ in token spending exiting 2028 (I am being Uber conservative - because it I apply the 3-5x pricing that Amy referenced, and apply that to capacity, the numbers become absurd) . It could get there, but the bar is high, especially given the dynamics above. I think the odds of a capex hiccup are a lot higher than what the market is pricing. The hyperscalers will be fine as they can just pause capex and digest what they have built out... The capex is still small relative to their earnings power. The long-term winners will be the companies that use cheap intelligence to rethink their existing businesses or build new ones, not the capex plays.

Obamacare was sold as expanding coverage while bending the cost curve and cutting typical family premiums by ~$2,500. It substantially expanded coverage—the uninsured rate fell from roughly 15% pre-full implementation to around 8% in recent years—but never achieved universal coverage. The premium-cut promise was not met; costs continued rising (individual-market sticker prices rose sharply for the unsubsidized due to guaranteed issue, essential benefits, and rating rules) at a faster rate than overall inflation. Democrats have now doubled down on the same failed ideo by pushing for this concept of Medicare for All. That's just more government involvement in the largest segment of the U.S. economy. When was the last time U.S. government involvement in any sector resulted in more efficiencies? Look at how ridiculously wasteful our defense spending is, and why new private companies like Anduril are able to build weapons at a much lower cost.

The recent drawdown in AI-capex-tied equities, and the open schadenfreude (as well as posts of remorse) that has accompanied it, got me thinking. You are never as good as you think you are when you are winning, or as bad as you think you are when you are losing. The roll of the dice plays a far larger role in investment and life outcomes than most people care to admit. Michael Mauboussin captures this dynamic brilliantly in *The Success Equation*, where he untangles the interplay of skill and luck across markets, sports, and business. His most counterintuitive insight is the “paradox of skill”: as any competitive field matures and the baseline level of talent, data, and analytical rigor rises, the gap between the best and the average participant narrows. Paradoxically, as skill becomes more uniform at the top, luck becomes a bigger determinant of the final outcome. Nowhere is this clearer than in investing. At the institutional level, hundreds of smart, relentless analysts are parsing the same filings, tracking the same alternative data, and building similar models. When everyone at the table is highly capable, small exogenous variables—a sudden macro regime shift, an unexpected regulatory headline, or the exact timing of a liquidity squeeze—can turn a mediocre thesis into a windfall, or a rigorous, high-probability bet into a painful drawdown. Looking backward, we hate randomness. We invent clean narratives about superior foresight during bull runs and fatal flaws during drawdowns so the uncertainty feels less threatening. Recognizing the true weight of luck doesn’t mean abandoning rigor—it means judging yourself by the quality of your process rather than the noise of a single outcome. It demands humility when your numbers look brilliant and resilience when the tape goes against you. Skill and effort matter enormously, and so does the roll of the dice. Both are true at once. And my final advice - while it's a good idea to let your winners run, take profits along the way. It's more important to live to fight another day.

Looking increasingly prescient. Consensus trades - when Everyone is positioned the same way - are dangerous...