Free Weights, Expensive Answers
Mira Murati just gave away a 975-billion-parameter model. Before you forward this to your board — read the fine print I read.
A CFO leaned across a table in SCBD last month and asked me whether his group should “build our own AI.”
Not use. Build.
I asked him one question back: “What would you put in it?”
He’s still thinking about it.
On Wednesday, that pause got a price tag.
What Murati actually shipped
Thinking Machines Lab — the startup the former OpenAI CTO founded last year — released Inkling.
The headline numbers:
975 billion parameters, mixture-of-experts — only ~41 billion fire per task
1-million-token context window
Trained on 45 trillion tokens: text, images, audio, video
Weights free on Hugging Face. Fine-tuning via Tinker, their customisation platform
Largest American open-weights model ever released. Nvidia’s Nemotron 3 Ultra held that crown at 550 billion.
And then the lab said something no frontier lab says.
In its own launch post: Inkling is not the strongest model available today. Open or closed.
That’s not humility.
That’s a business model.
They’re not selling capability. They’re selling ownership.
“Free” — a word your auditor should flag
Here’s the number nobody puts in the headline: two terabytes.
That’s the GPU memory Inkling needs at native precision — roughly eight Nvidia B300s, or sixteen H200s, per The Register. A quantised version halves it.
Sixteen H200s.
Try sliding that past an audit committee in Jakarta. Or Coimbatore. In IDR or INR, that line item has its own gravity.
Free weights are free the way a puppy is free.
The download costs nothing. Everything after the download costs everything.
The three questions before any board funds a build
I’ve started asking these in every AI steering-committee meeting. Your proprietary knowledge must survive all three:
#The questionWhere builds die1Does the base model already know this?If GPT knows your industry cold, you’re fine-tuning air2Does your knowledge compound?One-time knowledge is a prompt. Compounding knowledge is an asset3Are you leaking it by renting?Satya Nadella’s warning: closed-model customers pay twice — once in fees, once in the expertise handed over inside every prompt
Bridgewater passes all three. The hedge fund fine-tuned Alibaba’s Qwen on its own financial reasoning via Tinker — and reports 84.7% on financial reasoning evals, beating leading proprietary models at a fraction of the cost.
One word of caution before that number reaches your board pack: reports. It’s the companies’ own evaluation. No independent verification yet. Caveat it, or don’t cite it.
Most firms I meet fail Question 2.
Comfortably.
Why this matters more in our corridor
The India–Indonesia stack has spent two years renting intelligence and calling it strategy.
Meanwhile the Western open ecosystem thinned out after Meta’s Llama 4 stumble pushed it proprietary — leaving Chinese models as the default alternative. For an Indonesian bank answering to OJK, or an Indian NBFC, that has always been the awkward slide in the data-sovereignty deck.
Inkling makes that slide less awkward.
It does not make the build cheaper.
So here’s the argument in one line:
Open weights don’t cut your AI bill. They convert a rental expense into an owned asset — and most organisations have nothing worth capitalising.
That CFO in SCBD hasn’t answered my question yet.
But he’s asking the right one now — not “should we build?” but “what do we know that nobody else does?”
That question has never needed a GPU.
What would you put in yours? Hit reply — I read every one.
Lift as you Rise.
CA Loganathan Anandan, FCA · CISA · CDPSE · CFE