Two rounds, one label
The larger deal made a storage company Israel’s most valuable private technology firm: a billion dollars raised at a thirty-billion valuation, half of it buying out early holders, on recurring revenue said to have grown tenfold in two years by keeping accelerator clusters supplied with data. The chip company that backs it put its chief executive in the launch video to say that even the fastest processors hit severe data bottlenecks without this software underneath them.
The smaller deal was faster-moving: a one-year-old enterprise agent vendor at two billion, planning to nearly triple its headcount, on self-reported results of sixty percent faster handling times and containment above eighty percent — figures published without a methodology, and worth reading as a claim.
Both were described as infrastructure “beneath the model”. Both are real businesses solving real problems. And they are solving different problems, which is where an enterprise buyer needs to slow down.
Feeding the GPU is not governing the truth
The storage story is about throughput. Training and large-scale inference are bounded by how fast data reaches the accelerator; a platform that removes that bottleneck earns its valuation from cloud providers and model builders whose economics are dominated by idle GPUs. It is a genuine category, and for organisations running serious training workloads it is a genuine purchase.
The enterprise problem is different. An established organisation’s constraint is rarely how fast bytes reach a GPU. It is whether the data those bytes represent is correct, classified, permissioned and traceable — whether an assistant grounded in it can be trusted, and whether an agent acting on it can be audited. That is the warehouse’s job, the catalog’s job, the job of decades of unglamorous information management. No amount of throughput substitutes for it.
The confusion arises because “AI data platform” is now used for both, and because the throughput vendors, flush with capital, are extending upward into data management features — cataloguing, orchestration, a database of their own. That is a natural strategy for them. It does not mean the system of record should move.
The agent vendor’s promise, read carefully
The agent company’s chief executive said the decisive factor for enterprises this year would be “who can deliver deep integrations across complex infrastructures”. That is exactly right, and it is a claim about the same layer: the connective tissue between an organisation’s systems of record and the AI acting on them. Deep integration is the whole difficulty of enterprise AI, and a vendor promising it should be evaluated on precisely that — the write-back into the CRM, the permissions inherited from the warehouse, the audit trail in the ticketing system — not on containment rates measured somewhere else.
The self-reported figures deserve the usual treatment. Sixty percent faster than what baseline, on which workflows, measured by whom? Containment above eighty percent of what volume, with what escalation rate and what error rate among the contained? Those are answerable questions, and a serious vendor will answer them on your data during a pilot. The number that goes in the business case is the one from the pilot.
A way to sort the pitches
When a platform is presented as “for AI”, three questions place it.
- Does it hold the truth, or serve it? If your customer master, your ledger or your governed warehouse would move into it, it is a system-of-record decision and should be evaluated as one, with all the migration, lock-in and exit questions that implies. If it caches, indexes or streams data from the governed platform to models, it is a serving tier, and the questions are throughput, cost and — as with any copy of governed data — classification and retention.
- Where does governance live afterwards? The right answer for most organisations is “where it lives now”, extended to the AI tier through governed access, not relocated into whichever vendor raised most recently.
- What was measured, and on whose workload? Storage throughput on a reference cluster and agent containment on a reference process are both fine as vendor benchmarks and both meaningless as forecasts until re-run on yours.
The capital is right about the layer
Investors are correct that the value is beneath the model, in the layer between an organisation’s data and the AI that uses it. Where the enterprise should disagree is about who owns that layer. The vendors would like it to be them. For an established organisation with a governed data platform, the better answer is that the layer already exists, it is yours, and the purchases — fast storage, agent platforms, retrieval tiers — should plug into it rather than replace it. The label on the box matters far less than that.
