AINEWS
Together AI's $240M IBM Cloud deal shows the new AI hardware buyer: the cloud companies. Now they're starting to rent the boxes back.
Together AI just signed a roughly $240 million deal to run its next-generation training fleet on Nvidia HGX B300 systems. It's buying them through IBM Cloud, not from Dell, HPE, or any of the traditional server vendors (The Register).
The deal lands in Q1 2027 and is small in the context of hyperscaler AI capex projected at $725-800B for 2026 across Microsoft, Google, Meta, and Amazon (ValueAdd). But it is the cleanest signal yet that the biggest customers for AI hardware are no longer the enterprises who used to buy the boxes. They are the cloud companies. And they are starting to rent the boxes back.
The first lever is vertical integration through custom chips. Google TPU, Amazon Trainium, Meta MTIA, and Microsoft Maia are each shipping a new generation this year, with Broadcom's ASIC partnerships anchoring the custom-silicon roadmaps at multiple hyperscalers (Tom's Hardware, Hashrate Index). When your supplier designs its own accelerator, the merchant accelerator market narrows for everyone else.
The second lever is allocation. Nvidia's HGX B300 is the most sought-after AI server platform of the cycle, and hyperscalers consume the bulk of every shipment. An enterprise IT buyer walking into Dell or HPE in 2026 is buying from a reseller whose upstream allocation is already spoken for. HPCwire framed the current moment bluntly: the cloud already ate your hardware lunch (HPCwire).
For enterprise IT, the practical consequence is a narrowing of the menu. Three options remain: rent from a hyperscaler, rent from a hyperscaler-adjacent provider like IBM Cloud or CoreWeave, or wait in a queue behind both. The merchant server OEMs are still selling boxes, but their allocation is thinner and their pricing power is lower. AI Capital Advisory's infrastructure analysis pegs supplier-side concentration as the dominant force in 2026 deal terms (Al Capital Advisory).
The Together/IBM structure is the visible shape of that concentration. Together AI is itself an AI cloud provider, but it has chosen to rent Nvidia's flagship platform from a hyperscaler-scale operator rather than build and buy the rack-scale infrastructure itself. The same logic applies to any enterprise that needs next-generation AI compute at scale: the path of least resistance runs through the same three or four counterparties.
The $725-800B capex figure is an analyst aggregation from CFA and venture research, not audited company guidance. It is directional, not exact, and it represents gross spend before any normalization for cancellations or delays. The Together/IBM dollar figure is reported but not yet audited. The custom-silicon programs are real, but the share of hyperscaler training and inference that runs on first-party silicon versus merchant Nvidia varies by company and is not publicly broken out.
The Q1 2027 Together/IBM deployment will be the first test of whether a non-hyperscaler AI lab can reliably rent next-generation Nvidia capacity at scale. If it works, expect more deals of this shape: AI-native companies buying compute through cloud operators rather than building their own data centers. If it slips, expect the queue behind the hyperscalers to harden further.
Either outcome, the buyer-side leverage in 2026 sits with a small number of hyperscalers. The traditional server vendors are still in the market, but they are no longer setting its terms.