The Compute Crunch Coming for Builders

Tom Tunguz's analysis of the emerging compute crisis lays out the math plainly: demand for training and inference is growing exponentially while the supply of chips remains constrained by semiconductor manufacturing realities. This means the cost of GPU hours won't drop as fast as founders hoped. Your Series A budget that seemed reasonable six months ago? It's now funding 40% less compute.

Why does this matter? Because it changes everything about how you build. If you're betting on training massive proprietary models from scratch, you're now competing with unlimited budgets from the hyperscalers. If you're planning to fine-tune on proprietary data at scale, the math just got harder. The smart move? Build on top of existing models through APIs and carefully targeted optimization rather than capital-intensive training runs. The constraint forces clarity—you have to know exactly what your model needs to do better, not just chase metrics.

This also explains why OpenAI's release of GPT-Rosalind matters beyond life sciences. Vertical-specific models reduce the compute needed to solve domain problems. Instead of adapting a general model, you get something pre-optimized for your use case. For founders in biotech, drug discovery, or any specialized domain, this is the playbook: find (or build) the right foundation model for your vertical, then layer your differentiation on top with lighter-weight techniques.

The broader pattern is consolidation disguised as specialization. We're moving from a world where anyone with access to cloud GPUs could train a competitive model, to a world where only those with significant capital or distribution can afford raw training costs. This pushes smaller builders toward:

1. Inference optimization – getting more value from cheaper inference via distillation, quantization, and efficient architectures. (See the MacMind project running neural nets on a 1989 Macintosh—constraints breed creativity.)

2. Vertical focus – build for a specific industry where domain knowledge compounds your advantage beyond raw model size.

3. Infrastructure play – the real money might be in making AI cheaper to run, not training it.

Enterprise buyers are catching onto this too. The framework shift toward "AI as operating layer" rather than point solutions reflects this reality. If compute is the constraint, then embedding AI into existing workflows beats bolting on new tools. For B2B founders, this means thinking about integration and infrastructure, not just capability.

The irony is that constraint-driven engineering produces better products. When you can't brute-force your way to answers with unlimited compute, you actually have to be clever about algorithm design, data efficiency, and user experience. The companies that thrive over the next two years will be those that treat compute scarcity as a design requirement, not a problem to throw money at.