SAN FRANCISCO, 14 AUG 2026 — Databricks has closed US$5 billion at a US$190 billion valuation, its second round this year. On the same day, chief executive Ali Ghodsi said artificial general intelligence has already arrived.
Read what he actually said, though, because it is close to the opposite of what the headline suggests.
The round
The money is pointed at three products, and the list is more revealing than the valuation. Lakebase is a serverless Postgres database for software that agents build, already at a US$100 million run-rate. Genie gives an assistant access to enterprise context. Unity AI Gateway routes workloads across models and controls what they cost.
Every one of them is plumbing, connecting a model to data an organisation already holds. The round buys no model at all.
What Ghodsi actually claimed
He defines AGI with deliberate modesty: "If a system can perform intellectual tasks humans perform and is smarter than most people most of the time, it meets the basic definition of AGI."
He then rules out the reading most people will assume. Asked about superintelligence, he is explicit: "If that is your definition, then of course it is not here."
And the sentence that carries the argument: "The world remains largely unchanged, except that token spending is rising."
The claim is deflationary, wearing a triumphalist headline. Ghodsi argues that the AGI threshold has already been crossed, and that it changed remarkably little. The binding constraint turns out to be enterprise context — the unglamorous problem of a model not knowing what your company knows.
The position is coherent, and it is also his book
Both things are true and neither cancels the other.
The argument holds up on its own terms. If capable models are widely available and the observable effect is mostly a larger inference bill, then the bottleneck has moved downstream, into data access and retrieval quality and the messy business of connecting a model to systems designed before it existed.
It also happens to be the exact thesis that makes Databricks worth US$190 billion. A company selling the layer between models and enterprise data has an obvious interest in the world believing that layer is where the value sits.
He is not necessarily wrong. But the claim is an argument, not testimony, and the strongest evidence for it is not what he says — it is what the round is being spent on. Nobody raises five billion dollars for context plumbing in a world where the model is the product.
The number underneath
Strip the AGI framing away and the revenue is what remains interesting. A US$7 billion run-rate, growing more than 80 per cent.
At that scale, 80 per cent growth is unusual enough to deserve its own attention. It is also the number that makes the valuation make sense: roughly 27 times run-rate revenue. That multiple is aggressive, but not detached in a market where the comparables are growing much more slowly.
Databricks disclosed no profitability figure. Nobody expects a company growing 80 per cent while funding three new product lines to be profitable, but its absence should be recorded rather than assumed.
The queue is the story under the story
That a company with a US$7 billion run-rate can raise its second private round of the year at US$190 billion, while calling a listing unlikely, says a lot about how capital works in this cycle.
Databricks is comfortably large enough to be public. It has the revenue, the growth and the investor base — Blackstone, T. Rowe Price and Point72 are not venture funds waiting for an exit, they are the institutions that normally buy on the public market. They are simply buying earlier.
The effect is that the growth happens where retail investors cannot reach it. A generation ago a company at this scale would have listed years back, and the appreciation from US$50 billion to US$190 billion would have accrued to anyone with a brokerage account. Now it accrues to whoever was allowed into the round.
Ghodsi's framing of the timing is precise and worth quoting for what it implies: a listing before Anthropic or OpenAI go public is "very unlikely". Readiness is not the issue. Ghodsi wants the market to price the frontier labs first, so that everything below them has a reference point.
It also means the queue has a head, and the head has not moved. Anthropic is reported to be preparing for a possible autumn debut and OpenAI has completed a US$7 billion share sale ahead of a potential listing of its own. Whichever goes first sets the multiple everyone else is measured against — including this one.
What it means for buyers in the region
Three practical reads for anyone in ASEAN evaluating this layer of the stack.
First, the products being funded address the actual complaint from the field. Enterprises here have spent two years discovering that a capable model plus inaccessible internal data equals a demo. Lakebase, Genie and a gateway that controls model spend are direct answers to that, and the fact that a US$190 billion company is betting its round on them tells you where the difficulty lies.
Second, do not overlook the cost-control piece. Unity AI Gateway exists because organisations have discovered that model spend is hard to see and harder to cap. That is the same problem we described in the LiteLLM breach, from a different angle — provider keys sitting in environments with generous quotas and no ceiling. Governance of model spend is becoming its own product category, and buying it from a vendor is one option among several.
Third, the unspoken risk: at 27 times revenue, this price assumes the plumbing layer stays valuable. If model providers absorb context and routing into their own platforms — which several are visibly attempting — the plumbing thesis weakens considerably.
What to watch
Whether the AGI framing gets repeated without its qualifiers. Ghodsi said the world remains largely unchanged; that half travels less well than the first half, and within a week the quote will circulate without it.
Lakebase's US$100 million run-rate is the small number in this announcement, and the one that tests the thesis. A database for agent-written software will either find a market quickly or reveal that agents are not yet building much that needs one.
Finally, the IPO comment. Saying a listing is unlikely before Anthropic or OpenAI go public is a way of describing a queue, and queues move.
