When the Model Is Free, the Hard Part Gets Harder

The cloud API tax on AI applications is becoming optional. MIT-licensed models at frontier scale, like GLM-5 (reported at 744B parameters), and open mixture-of-experts models like DeepSeek-V3 (with context windows in the 128K range) can now run on infrastructure you control, with tooling like Ollama making deployment routine rather than heroic. Whether these models match the best proprietary systems on every task is still contested, and benchmark parity fluctuates by workload. The economics now matter more than whatever gap remains.

Cost savings are the obvious consequence: lower bills, faster experimentation, weaker lock-in. What open weights do to differentiation matters more, and the effect is harsher than most builders expect.

APIs already commoditized access. Open weights commoditize the price.

It is tempting to say the API tax was a moat. It was not, at least not for application companies. Metered APIs lowered the barrier to entry so far that anyone could ship an AI product without an ML ops team, which is exactly why wrapper startups proliferated. Access to frontier capability stopped being a differentiator the day it became a credit card form.

What API pricing did do was set a floor. Everyone paid roughly the same per token, unit economics looked similar across competitors, and the vendor captured the margin. Permissively licensed open weights remove that floor. Any inference provider can serve the same model, and competition among them pushes serving prices toward hardware cost. Teams that build their own serving stacks can push further still, at the price of real operational complexity.

Accessibility already happened. The real shift is that the last shared cost structure disappears, and with it the last cover for products that were never differentiated in the first place.

Where the advantage actually goes

If any competitor can serve the same weights at near-commodity prices, advantage migrates to inputs the model cannot supply. "Data and domain expertise are the new moat" is an old refrain in tech strategy, so it is worth being specific about what is different this cycle:

• Proprietary data you can actually use. The corpus you ground or fine-tune with is not a commodity, and for the first time you can adapt genuine frontier-scale weights rather than a restricted or distilled checkpoint behind a vendor's fine-tuning API.
• Domain evaluation. Knowing which failure modes matter in your vertical, and building the harness that catches them, becomes the scarce skill. Generic benchmarks say little about your edge cases, and with open weights you can test against the exact artifact you will ship.
• Fine-tuning craftsmanship. Permissive licenses mean adaptation is gated by skill, not by a vendor's terms. Teams that can do this well pull ahead of teams that can only prompt.
• Inference economics as a design variable. Quantization, batching, hardware selection, and serving architecture become direct margin levers instead of a fixed per-token fee. Knowing when this work is worth doing matters as much as doing it well.

What "running it yourself" really costs

The electricity-and-amortization framing understates the bill. Serving a 744B-parameter model means serious hardware, capacity planning, and people who understand serving infrastructure, a talent cost most application teams have never carried. In practice, "local" often means your own cluster or a rented GPU box, not a laptop, though smaller distillations and MoE architectures narrow the gap.

Most builders will not run their own metal at all. They will move to serverless inference providers hosting the same open weights, capturing much of the price benefit without the ops burden. That is a rational default, and it means the strategic question is less "cloud versus local" than "how far down the stack is it worth going for your margin structure." For low-volume or spiky workloads, proprietary APIs may still win on total cost. And proprietary labs may keep a capability edge on specific tasks, where paying the tax remains the right call.

The licensing detail

The MIT license does real work here. Previous open-weight releases often carried usage restrictions or custom licenses that made enterprise legal teams nervous, which pushed companies back toward vendor contracts even when the weights were technically available. A genuinely permissive license removes that forcing function, and it changes negotiating leverage even for companies that never migrate off a proprietary API.

The question that replaces "which API?"

If you are building on AI today, stop asking which model API to use. Ask what you own that a competitor with identical weights does not: the data you can legitimately accumulate, the evaluation harness that encodes your domain knowledge, and a clear-eyed decision about how much of the inference stack is worth owning versus renting. The model is becoming the substrate. The durable work sits everywhere else.