The Self-Hosting Math Changed While We Were Still Arguing About Sovereignty

Global AI spending is forecast to hit $2.59 trillion by the end of the year, up 47% from 2025, with infrastructure the single largest line in that number. Underneath most of it sits one decision, repeated per workload: rent the serving or own it. That was always a portfolio question rather than a doctrinal one, and the portfolio was drawn by economics.

Those economics moved this year. Most enterprise architecture is still drawn against where they used to sit.

What Gated Self-Hosting Actually Is

It was never sovereignty. Sovereignty is legitimate. Data residency rules are real, and weights you control sit inside infrastructure you own. The gate was cost at frontier scale. Qwen3.8-Max is 2.4 trillion parameters. Usable throughput from something that size means a multi-node buildout, GPU procurement, cluster management, ongoing inference tuning, and round-the-clock support for a system that can't go down. Cross that threshold, and you haven't self-hosted anything. You've started a model-serving business.

That hasn't changed. Frontier scale just stopped being the price of frontier-class capability.

Intelligence Per Dollar Is Falling Off a Cliff

DeepSeek-V4-Flash runs 284B total parameters with 13B active and a one-million-token context window, and it serves on a single four-GPU node. Qwen3.8-Flash-Next pairs a 125B model with a 51B N-gram embedding table and activates 6B parameters per token; quantized builds are already running on two 64GB cards. Both are open-weight. Both land near where the closed frontier sat a few months ago on the work most enterprises actually run at volume: extraction, triage, code review, document analysis, routine agentic loops.

A year ago, hosting a model with frontier-class capability meant a data center. Today, it means a node. The line separating sensible from delusional moved a long way, and most enterprise architecture decisions haven't caught up to where it landed.

Serverless Takes Out the Last Real Objection

The honest objection to self-hosting was never only capital cost; it was utilization. Owned GPUs bill whether or not you're using them, so the economics only work past a steady consumption floor. Below that floor, you are paying to keep silicon warm.

Serverless GPU platforms like Modal drop that floor dramatically. Per-second billing on your own weights, autoscaling to zero, no fleet you have to keep busy to justify. That covers the awkward middle, where the workload is large enough that per-token API pricing starts to hurt but too spiky to justify reserved capacity. Own the weights, rent the seconds, and revisit the buy decision when consumption stabilizes.

Post-Training Is Where This Gets Genuinely Interesting

The bigger shift isn't cheaper serving. It's what you can do to a model once you control the weights.

Most enterprises pay for firm-specific context the same way every time: stuff it into the prompt, pay for it on the way in, pay again on the way out, on every transaction, forever. Engram's work with Harvey is the clearest picture I've seen of the alternative. They built a synthetic law firm, 266 matters and 9,286 files across 100 million tokens, then trained an agent to study the whole workspace before answering anything, holding what it learned in a mix of text notes and weights. It runs on Qwen3.8-27B and answers at $0.13 per query against Opus 4.8 at $1.32, while scoring higher. Probed closed-book, with no tools and no notes, the model's knowledge of the firm climbed from 4.7% to 72.6% over the course of studying.

Their framing is the part worth stealing: studying shifts compute from inference time to training time. Where many queries hit the same corpus, paying once to internalize it beats re-reading it on every call. That's an amortization argument, and it strengthens the longer a workspace lives, which describes nearly all enterprise work.

Knowledge is the easy half. Capability is the half worth watching, and it runs on the same machinery the open labs converged on this year. The student generates its own rollouts inside an environment you define, and the teacher model grades what it produces token by token. Engram's recipe uses both halves, distillation on model outputs alongside reinforcement learning on problems the model writes for itself. Pull that thread far enough, and the frontier model stops being a runtime dependency and becomes a training input. You pay frontier prices once, during post-training, rather than on every inference call for the life of the system.

Where Governed Hosting Still Wins

A governed arrangement through AWS Bedrock or Databricks remains the right call for a real set of situations: workloads that genuinely need the current frontier and nothing below it, organizations with no ML engineering capacity to speak of, and volumes low enough that per-token pricing never becomes the largest line on the bill. Those are real conditions, and plenty of enterprises meet them. Far fewer meet them across their whole estate, which is how most of them are buying.

Capability was never a pure compute question. It's how infrastructure decisions get made, what security posture is in place, whether an evaluation framework exists to tell you the model is actually performing, and whether anyone is tuning it as usage evolves. Self-hosting neither creates that burden nor removes it. It makes the burden much cheaper to carry, which means the binding constraint is now your people rather than your GPUs.

The Math Worth Running

Can we afford to be a model-serving company, and do we actually need to be one? For a growing share of workloads, that answer has quietly turned to yes, and most budgets are still built against the old one.

What fraction of my traffic genuinely requires frontier intelligence, and what fraction is running on it out of habit? What is my steady-state consumption, and where does it cross the line where owning beats renting? What would it be worth to bake my firm's context and my firm's tasks into weights I control, instead of re-paying for both on every single call?

Access to a frontier model's weights is still not the same thing as access to frontier capability. But the weights got cheap enough to serve and malleable enough to shape, and capability turned into something you can build rather than only something you can buy.