Theses›Datacentre & AI compute›Inference infrastructure

Inference infrastructure

Post-training becomes inseparable from inference — the data flywheel as the moat on open weights

  • Bull
  • Closing
  • $23B by 2035

The callEnterprises and states want to own their AI on open weights, but the retraining-loop moat is now consensus, priced at 17-22x revenue, and may sit with the app owner.

D1D2D3R1R2R3C1foragainst D1D2D3R1R2R3C1foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the drivers.
  1. D1Open weights match the best closed models (Kimi K3 in July 2026), so enterprises and states can own their intelligence rather than rent it.
  2. D2Lab and hyperscaler APIs are generic and don't learn from a customer's traffic, and sovereign and regulated buyers can't send their data to them.
  3. D3Government money for sovereign deployments: about $80B of sovereign cloud spend in 2026 and up to 7 EuroHPC AI Gigafactory sites.
  4. R1The moat is damaged by events: OpenAI wound down fine-tuning, Cursor does post-training in-house, Fireworks and Together make the identical pitch, token prices keep falling.
  5. R2If the data flywheel lives with the app owner or the lab, the capturable independent slice is the low end, about $1.5-2B.
  6. R3The cohort is priced at 17-22x forward revenue (Fireworks $17.5B, Baseten $13B, Together $8.3B), and even Scale AI became a feature when Meta bought about 49%.
  7. C1EuroHPC AI Gigafactory selection: up to 7 sovereign sites buying post-train-and-serve loops on their own data.early 2027
Rests onthe beliefs underneath
Supported by · confidence
What would prove itdated tests, open and resolved
20272028today1234predictionswatch 20272028today1234predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1By end-2027 the 'own your intelligence on open weights + post-training loop' pitch does NOT produce a defensible moat for any pure-play inference-serving vendor — differentiation stays execution/logos, gross-margin and multiple stay pressured by token-price deflation and identical competitor offerings — UNLESS one vendor demonstrably locks customers via proprietary data/eval lock-in rather than price/latency.Dec 2027 · open
  2. 2The EuroHPC AI Gigafactory selection, and which vendors the sovereign sites buy loops from.Dec 2027 · open
  3. 3No standalone serving vendor with a structural moat from open weights plus post-training; differentiation still execution and logos.Dec 2027 · open
  4. 4One vendor showing durable data and evaluation lock-in, or a margin washout across serving that cheapens entry prices.Dec 2027 · open
Expressionslisted, private, technologies
Private · 8 private companies, 7 of them in the US; none in continental Europe.

Baseten · Callosum · Fireworks AI · Together AI · Modal · fal · Replicate · Prime Intellect

Technologies
Who has already bet on itprivate rounds, from public sources

$4.7B raised by 7 companies, $4B of it in the last three years. Wave under way: Large rounds in the last three years; the category is funded and getting crowded.

Backed by a top-tier venture firm: Together AI, Modal, Fireworks AI, Prime Intellect, fal, Replicate. Backed by a strategic investor: Together AI, Fireworks AI, Prime Intellect, fal, Replicate.

What changedthe record, newest first
  • 26 Jul 2026Demand slice stronger, moat slice weaker, and the trade is priced at 17-22x forward revenue.
  • 20 Jul 2026The middle-power AI play is not building a national model, it is post-training someone else's free one on infrastructure you govern. Whoever sells that loop has a new customer class: governments.
  • 20 Jul 2026Cheap closed models undercut the cost case for running your own. So the reason to post-train an open model was never price. It is owning the data loop, or owning the jurisdiction.