Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO

Full analysis

Modal just raised $750 million at a $16 billion valuation on the back of a simple idea: nobody's AI workload is a rectangle. Capacity contracts are rectangles, fixed chips for fixed years, but real usage is bursty and spiky. CEO Erik Bernhardsson built a $1 billion-a-year business by pooling thousands of customers so he can rent you a slice when you need it and nobody pays for idle silicon. That's the pitch. The more useful part of the conversation is what he said about where GPU prices, open-source models, and non-NVIDIA chips are heading. If he's right, the cost of running AI drops, and a lot of people who signed big multi-year chip contracts are going to feel sick.

This is a briefing, not a decision. Nothing here forces a move this week. But two claims in it, if they hold, change the bills every AI buyer pays in 2027: fine-tuning is dying, and GPU prices might fall rather than rise. Both are easy to act on later and cheap to watch now.

The Skeptic

Bernhardsson is talking his book, and you should price that in. Modal makes money when you rent elastic capacity instead of locking in your own, so of course he warns that locking in is dangerous. His "GPU prices might collapse" line conveniently makes his rent-by-the-slice model look smart and everyone else's reserved-contract model look reckless. Convenient. That said, the fear is real. If you signed a three-year deal at 2026 prices and the market softens, you're the one holding overpriced silicon. His own company signs those same 3-year forward deals, so he's not immune. And "$1B ARR" at 30% month-over-month sandbox growth is the kind of number that only exists while the vibe-coding boom keeps paying.

The Open-Source Advocate

The line that should stop every AI buyer cold: "Fine-tuning is no longer something you have to do." Bernhardsson says open models like Kimi K3, DeepSeek 4.1, and GLM 5.3 are now close enough to frontier models that you just swap them in at a fraction of the cost. Fine-tuning means taking a base model and training it further on your own data to specialize it. For two years that was the standard advice for anyone building a serious product. If that work is now optional because the open models are good enough off the shelf, that's a direct cut to your build cost and your vendor dependence. It also means the pricing power of the closed labs on everyday workloads is eroding from underneath.

The Compute Pragmatist

Bernhardsson has predicted "the year of non-NVIDIA accelerators" every year for a decade and been wrong every time. He admits it. So take 2027 with salt. But the mechanism he describes is new: splitting one request across two kinds of chips. NVIDIA handles the heavy first pass (reading your prompt), and a specialist chip like Cerebras handles the token-by-token generation after. That's a real architecture, not a benchmark. NVIDIA keeps winning because its software ecosystem is mature and its competitors' tools are still painful to use. AMD and Google's TPUs post good benchmark numbers but the tools around them are still painful. Until that cracks, "non-NVIDIA" stays a line on a slide. Watch AMD's software, not AMD's chips.

The Builder

The quiet story is sandboxes. A sandbox is a throwaway cloud computer where an AI agent can run the code it just wrote without touching anything real. This went from near-dead in 2023 to a third of Modal's revenue, with median lifetime of about 8 seconds. That tells you where the agent economy actually spends money: not on clever prompts, on running generated code safely at scale. If you're building agents that write and execute code, this is now a first-class cost and latency line, not an afterthought. And the Hugging Face incident is the warning. An app running on Modal got exploited as a springboard into Hugging Face's systems. The infrastructure held; the customer app didn't. Agent-generated code running live is a new attack surface, and most teams haven't priced it.

Where they disagree

The real split is on GPU prices. The Compute Pragmatist and the Skeptic look at the same forward contracts and see opposite risks. If demand keeps climbing, scarcity wins and the people who locked in cheap look brilliant. If the open-model shift plus non-NVIDIA chips drops the cost of inference, the locked-in players are holding 2026-priced silicon in a 2028 market. Bernhardsson bets on the second, which happens to be the only world where renting beats owning.

The second disagreement: does the fine-tuning claim hold for YOUR workload? The Open-Source Advocate says swap in an open model and move on. The Skeptic notes that "almost as good" stops being good enough when your product lives or dies on the last 5% of quality on your specific task. Both are right depending on how commodity your use case is.

What it hinges on

Two beliefs, and you can test one of them cheaply right now. Is an open model "basically almost as good" at YOUR task? Run it. Take your current frontier-model workload, swap in Kimi K3 or DeepSeek 4.1, and measure the quality drop against the cost drop on your own traffic. If the gap is small, Bernhardsson is right and you've been overpaying. If it's large, your task isn't a commodity and fine-tuning still earns its keep. That's a one-week test and it settles the most expensive line in most AI budgets.

The GPU-price question you can't test, only position for. Don't sign a three-year chip commitment you can't walk back unless the price is good enough to survive a 30% market drop.

Prediction: By the end of 2027, at least one of the open-weight models in the Kimi, DeepSeek, GLM, or Llama families will rank within the top five on the LMArena text leaderboard, sitting above at least one flagship model from OpenAI, Anthropic, or Google.

Confidence: Medium. The gap has closed every cycle, but leaderboard timing is noisy.

Why: Bernhardsson says open models are already close enough that specialists are dropping fine-tuning and swapping them in for frontier models on real production work, and he runs the infrastructure that sees which models people actually deploy, so he's reading demand, not marketing. The direction has been one-way for two years: DeepSeek and Qwen variants have repeatedly landed near the top of public rankings within months of release, and the labs behind Kimi and GLM are shipping on the same cadence. For this to be wrong, the closed labs would need to open a quality gap wide enough that no open model cracks the top five all year, which would reverse the trend every recent release cycle has shown.

Revisit by 2027-12-31: We're right if an open-weight model from the Kimi, DeepSeek, GLM, or Llama families appears in the top five of the LMArena text leaderboard above at least one OpenAI, Anthropic, or Google flagship at any point in 2027. We're wrong if no open-weight model holds a top-five position above all three labs' flagships at any point that year.