The $4 trillion vendor problem you can't build around

Two things are happening at once, and they point in opposite directions. The money says AI is the biggest thing ever. The pricing says nobody knows what any of it is worth yet. You have to make roadmap calls in the middle of that. Let me catch you up.

The number that reframes everything

Here is the scale. SpaceX went public at a $1.77 trillion valuation, and with Anthropic and OpenAI pushing into the trillions, the three together land somewhere north of $4 trillion. A Pitchbook report puts it plainly: these exits will generate more value than every U.S. venture-backed exit since 2000. That's the era that gave us Google, Tesla, and Meta going public.

For context, the SEC counted just $70 billion in U.S. IPO proceeds all of last year. Uber's $84 billion IPO felt huge in 2019. It's less than 5% of what SpaceX just raised. Your vendors are becoming some of the largest companies on earth. That changes your negotiating position, and not in your favor.

Prices are still all over the map

While valuations climb, the thing you actually buy keeps swinging. Benedict Evans says only two things are certain about token prices: we're in a supply crunch, and it's unstable. Inference runs at 40 to 50% gross margins today, but that doesn't count training the next model, which still costs far more than revenue.

He makes the open question sharp. Do frontier labs end up with real pricing power, or do they become low-margin commodity infrastructure, like the cellular carriers who moved a trillion dollars of data and watched everyone up the stack capture the value? Evans thinks the signs point to commodity. That's good news for your bill and bad news if you're betting on one vendor staying special.

The bottleneck moved, and so did the money

Watch where the smart money went. Nvidia's stock fell 15% since May even as revenue grew, while Micron nearly tripled. The GPU shortage eased. Memory became the new choke point, with DRAM spot prices up tenfold in a year. As Ornn's Wayne Nelms put it, "Everyone wants to make their own silicon, but no one is making their own DRAM."

Google, Amazon, Microsoft, and OpenAI all built custom chips that are good enough to drag compute prices down. The price of an H100 hour peaked near $3.20 in May and slid after. So the input you depend on is getting cheaper and more competitive at the same moment your suppliers are getting richer. Both things are true.

Cheaper tokens might break the whole model

Here's the tension that should sit in the back of your head. Sequoia's David Cahn now pegs 2026 AI infrastructure spend at $1.5 trillion, which means the industry has to earn $3 trillion to pay it back. Anthropic is thought to be at $60 billion ARR, OpenAI around $20 billion. There's a real gap.

Apollo economist Torsten Slok flags the trap. Organizations keep shifting to cheaper open-weight models, often Chinese, and token prices keep falling. OpenAI's newest model is 54% more token-efficient on coding, per Sam Altman. Great for your budget. Rough for a token factory betting you'll spend more. Slok warns that if the hyperscalers miss their 2028 cash-flow targets, the fallout could tip the S&P into a correction.

The deep cut

The money is flooding in so fast that founders barely have to work for it. Lyzr let its own AI agent run a $100 million raise, pulling $400 million in interest without a single Sand Hill Road coffee. Mercor is in talks at a $20 billion valuation, double its October mark, after crossing $2 billion in run rate in four months.

That frothy capital is the thing to price into your build-vs-buy calls. It means new vendors will keep appearing cheaper and hungrier, so don't lock into a five-year contract on today's model. It also means the field will consolidate, so don't wire your product so tightly to one API that you can't swap it. Build a routing layer. Keep a cheaper open-weight fallback tested and ready. Treat any single model as replaceable infrastructure, because the economics say it is.

Three questions for your team

1. If our main model vendor doubled prices or got acquired next quarter, how many days would it take us to switch, and have we actually tested the fallback?
2. Which of our AI features has an ROI we could defend to a CFO today, and which are we running on faith because tokens felt cheap?
3. Are we architected to route across models based on cost and task, or have we hard-wired one API into the product where a swap would mean a rewrite?