The New Enterprise Battle Over Who Owns the Model

Full analysis

The story underneath all five news items is the same one: the enterprise AI question has stopped being "OpenAI or Anthropic?" and become "who holds the weights, and who touches your data?" Thinking Machines Lab's Inkling (a 975B-parameter mixture-of-experts model where only 41B fire per query) plus its Tinker fine-tuning API is the purest expression of that pitch — tune on your own infra, no leakage to the lab. Microsoft is running the same play from the other direction, telling its own sales reps to sell against Claude with its cheaper MAI models. This is a Type 1, hard-to-reverse question for anyone signing a multi-year model contract, and a Type 2, cheap-to-test question for anyone just adding a fine-tuning pipeline to one workload. The forcing function is real: Anthropic's fall IPO and ballooning token bills are pushing buyers to re-examine lock-in now.

The Skeptic — The load-bearing claim here is that fine-tuning your own open-weight model is a moat. Simon Smith already torched it: the real cost isn't the tune, it's the forever — re-curating data for new edge cases, babysitting drift, re-tuning every time the base model updates, and watching a bigger general model with a scrap of context lap your hard work six months later. Inkling ranks 19th globally at a 41 intelligence index, ~20 points behind Fable 5. You're volunteering for a maintenance treadmill to run a mid-tier model. For the PM: "own your model" often means "own a second full-time engineering project that a frontier API would have made obsolete anyway."

The Researcher — The genuinely novel thing isn't the benchmarks, it's the provenance. Jack Morris's claim — Inkling is trained without distilling from OpenAI or Anthropic, unlike Kimi, GLM, Qwen, Nemotron — matters because a fully independent tech stack means independent failure modes and no legal contamination from a competitor's outputs. But read the correction: there was a Kimi K2.5 synthetic-data bootstrap, so "from scratch" is doing some marketing work. The number I'd actually bank on is token efficiency: ~2/3 the tokens per task versus Kimi 2.6 and DeepSeek V4 Pro. That's a real, measurable architectural win. For the PM: fewer tokens to finish the same job means a smaller bill, and that's verifiable, unlike "our fine-tune is a moat."

The Open-Source Advocate — This is the good timeline. An open-weight model at 46% on Humanity's Last Exam with tools, a 1M-token context, native multimodal, and a real fine-tuning API — you can run the whole thing on your own hardware. The GKE security-blueprint and model-migration pain that Google Cloud keeps publishing about? That pain is why holding your own weights is attractive: no forced migration when a vendor deprecates a model out from under you. But be honest about the license and the reproducibility before you build on it. For the PM: open weights mean the rug can't get pulled — the model you shipped on is the model you keep.

The Compute Pragmatist — Follow the chips and the money and the story tells itself. Apple can't ship its Baltra server chip, is renting Google Cloud NVIDIA capacity, and is chip-shopping in a market where NVIDIA just bought Groq for $20B. Cursor wants a "significant compute advantage by 2027." Everyone is racing to own silicon because inference at scale is where margins live or die. Inkling's 41B active params (not 975B) per pass is the whole point — MoE means you pay for a fraction of the model each query. For the PM: the sparse-activation trick is why a giant model can be cheap to run, and token efficiency compounds that — it's the difference between a sustainable inference bill and a runaway one.

The Builder — What ships Tuesday? Not Inkling as your primary model — it's 19th and you'd be maintaining it. But Tinker as a pipeline for one narrow, high-value workflow where data can't leave your walls? That's a real weekend prototype. The honest move is what's already happening in the room: someone upgraded Claude Haiku to Sonnet and got "meaningful quality improvement" for near-zero effort. That's the baseline every fine-tuning project has to beat. Microsoft swapping Copilot functions to MAI for cost, then telling reps Claude is "slower and less accurate" in Office — that's an integration argument, not a capability one. For the PM: the cheapest win is usually a better prompt or a bigger off-the-shelf model, not a custom tune.

Where they split:

• Researcher vs. Skeptic on independence. The Researcher thinks a non-distilled, independent tech stack is a durable asset (clean IP, uncorrelated failures). The Skeptic thinks it's irrelevant if the model is 20 points behind frontier and you're stuck maintaining it — provenance doesn't pay the maintenance bill.
• Open-Source Advocate vs. Builder on ownership. The Advocate sees weight ownership as insurance against forced migration. The Builder points out that the room's actual win came from renting a better closed model (Haiku→Sonnet) with zero ops cost — the exact opposite of the ownership thesis.
• Compute Pragmatist vs. Skeptic on where the moat is. The Pragmatist says the real edge is inference economics — token efficiency and sparse activation. The Skeptic says none of that matters if a bigger general model plus a little context beats your tuned setup regardless of how cheaply it runs.

What it actually hinges on: two testable beliefs. First — does owned fine-tuning beat a big general model plus RAG on your specific task, and stay ahead across base-model updates? Second — is data sovereignty a hard requirement (regulated data, competitive exposure) or a nice-to-have you're paying a maintenance tax for? If sovereignty is a genuine constraint, Inkling/Tinker and Microsoft's Frontier Tuning are worth a bake-off. If it's not, the council leans hard toward the boring answer: rent the best closed model, add context, and revisit only when the token bill has a clear line to revenue that a cheaper open model would improve.

De-risk before committing: run the Skeptic's experiment explicitly — tune Inkling on your real task, then benchmark it against Claude/Gemini-plus-RAG on a held-out set, and re-run that comparison after the next base-model release to see if your tune's edge survives an upgrade. Model total cost of ownership including drift maintenance, not just per-token savings. Get license and indemnification terms in writing before any weight lives in production.

Prediction: By the end of Q1 2027 — one to two frontier release cycles out — no independent benchmark will show a general-purpose fine-tuned Inkling beating a top-3 closed model (Claude/Gemini/GPT) plus retrieval on a broad enterprise task; its wins will stay confined to narrow, data-sovereignty-constrained niches.

Confidence: Medium — mid-tier base plus the well-documented fine-tuning-drift trap.

Why: Inkling enters at 19th globally (index 41), roughly 20 points behind Fable 5 on Humanity's Last Exam, so the fine-tune has to close a large capability gap and then hold it. The mechanism working against it is the one Simon Smith names and the Google Cloud migration article confirms: base models keep improving, fine-tunes lose capabilities and require constant re-curation, and "a big general model with a bit of context" tends to catch up. For the tuned model to win broadly, TML's efficiency edge would have to outrun frontier capability gains between now and Q1 2027 — the less likely outcome given the pace of closed-model releases. The realistic win is narrow: regulated or competitively sensitive workloads where sovereignty is non-negotiable.

Revisit by 2027-03-31: We're right if fine-tuned-Inkling wins remain limited to sovereignty-driven niche deployments and no public eval shows it beating a top-3 closed model plus RAG on a general task. We're wrong if an independent benchmark shows a tuned open-weight Inkling matching or beating frontier closed models on a broad enterprise workload.