What each contract puts on paper

IBM says Together AI has signed a multi-year agreement worth 240 million dollars to run open-source model inference on NVIDIA HGX B300 systems hosted on IBM Cloud, with Spectrum-X Ethernet between them, and that the cluster should be available in the first quarter of 2027. That sentence carries one firm amount and one date. It carries no accelerator count, no megawatt figure, no site and no number of years for the multi-year term.[1]

The other contract, reported by The Decoder from Bloomberg's account, has Anthropic taking 191 megawatts at Riot Platforms' Rockdale site in Texas on a 20-year lease worth 9.1 billion dollars, rising to 16.1 billion dollars if both extension options are exercised. Riot puts up the building and supplies power connections, cooling and operations; Anthropic brings the servers and the chips. Riot had disclosed the contract a day earlier with its quarterly results, naming the tenant only as a leading frontier AI lab.[2]

Putting the two on one denominator

Let me put these on one denominator, because that is where they part company. From the Riot lease a reader can derive a price per megawatt-year, because the lease states an amount, a term of 20 years and 191 megawatts, and its two dated tranches say when that capacity is meant to exist. From the IBM agreement no equivalent figure can be derived, because the denominator — accelerators, racks or drawn power — was never published. The 240 million dollars is a signed amount attached to an unmeasured quantity.[1], [2]

The announcement does carry two numbers that look as though they would help, and neither closes the gap. Together AI is said to serve 400 trillion tokens a month, which describes today's workload rather than the cluster being bought, so without an accelerator count there is no way to ask how much of that traffic 240 million dollars pays for. IBM also repeats NVIDIA's claim that the B300 generation delivers 30 times more AI factory output than the one before it, with no workload, precision, latency target or power boundary attached; that makes it a vendor comparison, and taking it literally would move any capacity estimate by an order of magnitude. The likeliest reason for the silence is ordinary commercial confidentiality around unit pricing; another is that the deployment size is still being settled against NVIDIA's own delivery schedule.[1]

What would move each one up a rung?

There is a dated test here. If the cluster opens in the first quarter of 2027 as IBM states, the launch material will have to give a customer something to size against: an instance type, an accelerator count, or a published price per accelerator-hour. If that quarter closes with the service live and still no such figure, then 240 million dollars will have bought capacity that cannot be compared with any other cluster on the market, including the infrastructure Together AI serves its 400 trillion monthly tokens from today.[1]

Riot's side of the comparison has its own next rung, and that one is physical: 96 of the 191 megawatts are due online in December 2027 and the balance in June 2028. An earlier column here read Anthropic's Texas position through the interconnection queue, where 474 gigawatts of requests sat behind a far smaller amount of usable capacity. The Rockdale lease is that same constraint with a signed counterparty and two dates attached, which puts it a rung above a queue position and leaves it two energisation events away from a running machine.[2], [3]