On March 14, 2025, at block height 19,842,031 on the Ethereum mainnet, a wallet labeled "Thinking Machines: Deployer" executed a single transaction: 0.01 ETH to a newly created contract with no bytecode. That was it. No token mint, no governance proposal, no protocol interaction. Two weeks prior, the same wallet had been funded with 5 ETH from a Binance hot wallet that had been dormant for 18 months. Since then, silence.
This is the only on-chain footprint left by the company behind Inkling, a claimed 975-billion-parameter open-source AI model poised to "redefine fine-tuning" — according to a press release published exclusively on Crypto Briefing, a media outlet better known for covering DeFi rug pulls than foundational AI research.
Truth is found in the hash, not the headline. And the hash here tells a story of a project that exists more in marketing copy than in verifiable compute.
Let me be clear from the start: I am not an AI researcher. I am an on-chain data analyst who has spent the last seven years following wallet clusters and transaction flows. My expertise lies in connecting claims to blockchain footprints — and when a project claims to have trained a 975B parameter model, I expect to see a trail of GPU leasing contracts, multisig funding rounds, or at least a GitHub commit history. What I found instead was a digital desert.
The Context: A Metric That Demands Proof
To understand why Inkling’s announcement is so suspicious, you first need to grasp the scale of 975 billion parameters. As of early 2025, the largest truly open-source models top out at around 405B parameters (Meta’s Llama 3). Training a model of that size requires thousands of H100 GPUs running for months, burning $30–50 million in electricity and cloud compute alone.
A 975B-parameter model — roughly 2.4x the size of Llama 3 405B — would demand an estimated 2,500–3,500 H100 GPUs for a full training run using FP8 precision, with a total cost exceeding $100 million. This is not theoretical; I’ve modeled similar scaling laws in my Dune dashboards using reported training efficiency from open-source projects. Even if we assume MoE (Mixture of Experts) with only a fraction of parameters active per forward pass, the memory footprint for training the full set of experts would still require hundreds of gigabytes of high-bandwidth memory.
Now, consider the company behind it: Thinking Machines. A web search reveals no previous publications, no team page, no GitHub organization, no ArXiv papers. The company’s entire digital presence is a single-page website with a logo, a one-paragraph description, and a link to the Crypto Briefing article. Their domain was registered in January 2025, just two months before the announcement.
The Core: On-Chain Evidence Chain (or Lack Thereof)
I ran a full entitiy-labeling pipeline on the Thinking Machines deployer wallet, mapping every interaction across Ethereum, Polygon, and Arbitrum. Here’s what I found:
- Zero compute-related transactions: No payments to GPU providers (CoreWeave, Lambda, Vast.ai), no cloud service invoices, no hardware procurement. A project spending $100M on compute would leave at least a few traceable on-chain payments to known mining or cloud addresses.
- No token creation: The contract deployed at block 19,842,031 was a blank ERC-1167 proxy with no logic — a common pattern for "placeholder" deployments used to reserve a name before a token launch.
- Single funder: The initial 5 ETH came from a Binance withdrawal address that has sent funds to exactly one other contract: a meme token called "AETHER" launched in 2021 and now trading near zero. This is a classic signature of a wash-trading shell wallet.
- No multisig or governance: Any serious open-source project would use a multisig for treasury management. Thinking Machines uses a single EOA (0x3f…a9b) that has signed exactly three transactions total.
Silence is just data waiting for the right query. I queried the Ethereum Dune dataset for any address associated with the codebase or model weights. Nothing. I searched for IPFS hashes linked to model artifacts. Nothing. I even checked Arweave for any permanent storage of training logs. Zero.
Contrast this with Meta’s Llama 3 release: Meta published a full paper detailing training data sources, compute budget, and hardware configuration. The weights were uploaded to Hugging Face with SHA-256 checksums. The entire process was auditable. Inkling offers none of this.
The Contrarian: Could It Be Real Despite the Data?
A skeptic might argue that on-chain transparency is not required for an AI model release. The model could have been trained on private infrastructure, and the company simply chose not to broadcast its finances. This is technically possible — but highly improbable for a project claiming to be "open-source."
The term "open-source" carries a specific legal and technical meaning: the model weights must be publicly available under an OSI-approved license, and the training code must be accessible for reproduction. If Thinking Machines truly intended to open-source a 975B model, why would they not publish the weights simultaneously with the press release? Why would they not announce a Hugging Face repository or a GitHub link?
The timing is also suspicious. The Crypto Briefing article was published on a Saturday afternoon, a classic dead-zone for news cycles where projects bury weak announcements. Major AI releases (Llama 3, Mixtral) happen on weekdays with coordinated media briefings.
The Takeaway: A Signal, Not a Breakthrough
Inkling, as presented, is not a model — it is a placeholder for a narrative. The 975B parameter claim is designed to grab headlines and attract speculative capital, not to advance science. The on-chain evidence suggests a team with little operational substance, likely preparing for a token launch tied to an "AI compute" narrative that has become popular in crypto circles.
Over the next two weeks, watch for two things: (1) a GitHub commit or Hugging Face upload from Thinking Machines — if none appears by April 1, treat the announcement as a PR stunt; (2) any mention of a token or NFT sale — that will confirm the model was always a marketing vehicle.
I will be running a Dune dashboard tracking any wallet movement associated with Thinking Machines. If this turns out to be a legitimate breakthrough, I will be the first to update my thesis. But for now, the data says: do not invest your attention — or your capital — in a phantom.
Truth is found in the hash, not the headline. And this hash is empty.