Going long on compute takes more than chips: Silicon Valley's inside joke about structural shorts
Someone explained to the author that buying compute with equity alone isn't truly ‘going long’—you need debt leverage on chips too, or you're ‘structurally short’. It's a Silicon Valley inside joke.
The argument · tap a timestamp to hear it
Open Compute has overshadowed the legacy semiconductor trade show
Semicon West, which alternated between San Francisco and Arizona, is holding its last SF edition; the next moves to Arizona in March 2027. It's losing to OCP Global Summit, which dominates the same week with focus on inference-era infrastructure (cooling, networking). The author expects Semicon to eventually relocate to Texas, reflecting the geographic shift of US chip manufacturing away from Silicon Valley toward inland states.
Anthropic's $16 billion revenue fails to satisfy markets
During the author's visit, the Financial Times reported Anthropic's financials: $16 billion annualized revenue with excellent margins—numbers any industry would celebrate. Yet markets deemed it insufficient; AI stocks pivoted on the news. It's a snapshot of the disconnect between builders inside the conference and financial markets outside, anxious enough to stretch 2027 overcapacity worries into questions about whether 2028 demand can hold up.
Equipment capex forecasts hide a textbook bullwhip effect
Lam Research raised 2026 wafer fab equipment capex guidance from $140 billion to $150 billion and projects 2027 spending above $200 billion. Goldman Sachs forecasts $218 billion for 2027 and $281 billion for 2028. Yet at the same event, the author hears rumors that equipment suppliers are quietly freezing hires. Demand signals amplifying upstream while hiring throttles at the source: textbook bullwhip effect.
The jagged edge explains why startups must build custom models
Silicon Valley entrepreneurs repeatedly invoke ‘jagged edge’ to explain custom model development: today's large language models excel at some tasks while failing unpredictably at others, and the boundary is hard to predict. Step off the ‘golden path’ of well-trained tasks and performance cliffs. The author notes the same term describes Chinese open-source models—dominant in niche domains but with equally stark weak spots.
Not leveraging debt on chips is ‘structurally short’—a Silicon Valley inside joke
Smart investors believe every additional gigawatt of compute grows more valuable, so many prefer equity financing over debt to buy chips; debt default is a common company killer. But others push back: if equity-only chip acquisition can't match a rival's speed, raw compute scale crushes you anyway. Real ‘going long’ isn't just raising capital to buy chips—it's borrowing to add leverage on top and buy more. Anything else is ‘structurally shorting’ yourself. It's a running inside joke.
Startups burn through subscription subsidies to stay alive
Semianalysis research shows a $200/month Claude or ChatGPT subscription yields $2,000–$14,000 in equivalent API value. Cash-strapped startups accumulate multiple subscriptions, burning hard at token resets. Some route entire customer-facing products through these subsidized subscription tokens. The author sees this as obviously unsustainable, economically deformed business models masquerading as real companies.
AI subscription subsidies mimic the rideshare wars of 2015–16
The author recalls 2016 leaked Uber financials showing customers paid only 41% of actual trip costs—sometimes including extreme driver subsidies in China (three times the fare per ride). Today's AI subscription subsidies aren't yet that aggressive. But OpenAI's premium Pro 500+ tier and startup migration toward native API pricing show markets adjusting. Startups are rebuilding unit economics around true API costs and exploring load-balancing across multiple models.
Massive math API spending might fund automated research loops
OpenAI and Anthropic spent roughly $6.5 million in API calls on Navier-Stokes proofs and similar math problems. Even with ample compute, the scale seems extreme for a blog post. One theory: the spend primes Recursive Self-Improvement (RSI). Train models to reliably prove or disprove a hypothesis, then automate a sweep across human-generated hypotheses, verify them, and feed results back into training. By this logic, math spending isn't marketing but a cornerstone of coming R&D.
The true training-to-inference compute ratio is 75:25, not 50:50
The author's earlier 50:50 estimate assumed inference could absorb equal capacity. Infrastructure reality suggests otherwise. If RSI theory holds, the math resolves: actual split is around 75% training, 25% inference, with surplus training capacity feeding automated hypothesis verification. It explains why Anthropic and OpenAI will fund the coming torrent of new compute—not to brag, but to run RSI.
In their own words · checked verbatim
The numbers would be great for most any business. 16 billion revenue with excellent gross margins. But they were deemed insufficient by the market, throwing AI trade stock prices into turmoil.
The jagged edge is why a startup should build their own special models, things like world models, or models for handling special cases.
But put them outside of their golden path, and they supposedly fall far short. That is the jagged edge.
You should then be taking out debt on top of those chips so you can buy even more chips. If not, then you are not truly long. You are in fact, quote unquote, structurally short. I laughed when I heard that.
In 2016, Uber's leaked financials showed that its customers were only paying 41% of the actual costs of their trips.
If it's for me that these companies invest in solving math problems more because it is a critical part of their future R&D plans, and not for flexing and or marketing.
Rather than an even 50-50, it turns out to be more like 75-25 training to inference.
Silicon Valley still marches forward along the jagged edge. But falling off now somehow feels easier than before.
Figures
| Anthropic annualized revenue | $16 billion | 4:07 |
| 2027 year-end compute capacity target | 50 gigawatts | 4:07 |
| Lam Research 2026 WFE capex guidance | $150 billion | 5:16 |
| Lam Research 2027 WFE capex guidance | $200+ billion | 5:16 |
| Goldman Sachs 2027 WFE forecast | $218 billion | 5:16 |
| Goldman Sachs 2028 WFE forecast | $281 billion | 5:16 |
| $200/month AI subscription API equivalent value | $2,000–$14,000 | 9:26 |
| Math problem API costs (Navier-Stokes proofs) | $6.5 million | 13:34 |
| 2016 Uber customer payment as share of cost | 41% | 11:30 |
| Revised training-to-inference compute ratio | 75:25 | 16:41 |
Glossary
- Jagged edge
- The phenomenon where large language models excel at some tasks while failing unpredictably at others, with performance boundaries hard to anticipate.
- Recursive Self-Improvement (RSI)
- An automated process where AI models verify hypotheses and feed results back into training, enabling research scaling without human iteration.
- Bullwhip effect
- A supply-chain pattern where demand signals amplify as they move upstream, often triggering hiring freezes at the source despite strong forecasts.
- Golden path
- The domain of tasks where an AI model excels after thorough training; leaving it causes sharp, unpredictable performance drops.
- Structurally short
- Using only equity capital to buy compute capacity without debt leverage, colloquially viewed as betting against your own position.
How to listen
Investors and practitioners tracking the AI compute arms race, semiconductor equipment capex forecasts, and the real cash flow situation at AI startups.
The opening about the author's conference experience and trade show changes can be skipped—information density is low there.