Here's What's On Our Plate Today

• 🧪 The $25.5B bet on rented intelligence: legal AI doesn't own its brain.

• 🍪 OpenAI's data center chief exits, Instagram auto-cuts Reels, Taiwan indicts nine over server smuggling.

• ✅ Weekend To-Do: price your dependency, read Harvey's benchmark, run a substitution test.

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

The PC makers survived Intel. That is the optimistic case for legal AI.

• The prices: Harvey is in talks at $15.5B, and Legora is above $10B, roughly 44 and 67 to 80 times revenue. Neither has closed.

• Rented intelligence: both build the workflows, integrations, and security clearances around models licensed from OpenAI, Anthropic, and Google. Neither owns the part that thinks.

• The supplier moved in: Anthropic's January legal plugins cut Thomson Reuters' stock 15.83% in one session. In May, it shipped practice-area legal tools and a connector into Harvey.

• Harvey's own numbers: its internal legal test showed seven newer models beating the tuned version it originally measured. The advantage expired without anyone attacking it.

• Stakes: the real risk is not replacement. It is surviving while the money drifts toward the part nobody can substitute, and no product launch announces the day that starts.

Harvey, Legora & the risk of renting your intelligence

The modern economic system is a remarkable beacon of human ingenuity and of the trust that organizations extend to one another. It works only because everyone involved has agreed to a division of work, where many separate companies each build one piece of something complicated, like a laptop or a smartphone. Take the personal computer, where, for most of its history, the company whose name was on the lid did not make the processor inside it. Dell, HP, Lenovo, and Acer bought their chips from Intel and spent their own money on everything else: the case, the keyboard, the screen, the cooling, the warranty, and the sales teams that got those machines into corporate purchasing departments. The processor was the one piece they rented from somebody bigger, and the arrangement rested on a piece of trust nobody had written into a contract: that the chip maker would stay within the part of the job it had agreed to do. The assemblers were confident enough to advertise it, putting a sticker on the outside of their own machines to announce whose part was inside. What eventually went wrong was quieter than a betrayal, because Intel never had to build a laptop for the money in a laptop to drift toward the chip. Apple was the only company that refused the arrangement outright and designed its own processors, an enormously expensive way to solve the problem, and the only one that solved it completely.

The same division of work now sits at one of the most expensive corners of enterprise software, carrying the same unwritten trust. Harvey and Legora sell artificial intelligence software to law firms, and neither of them builds the intelligence. Both rent it from OpenAI, Anthropic, and Google, then build around it the parts they do own: the connections into a firm's document systems, the security clearances that get a tool past procurement, and the relationships with the partners who sign the contracts. The question underneath both businesses is whether that second set of things is theirs to keep, or whether it is work the companies supplying the intelligence will eventually decide to do themselves.

Whether that trust holds decides considerably more than who wins a venture round, because if it holds, a generation of companies built on other people's models has a real future, and if it fails, the turn arrives fastest for the law firms that moved earliest. Two funding conversations in August put a price on which way the market currently leans.

A price set in two conversations

On August 7, 2026, The Information reported that Harvey had opened talks to raise at least $500M at a valuation of $15.5B, a 40% step up from the $11B it was worth five months earlier. Six days later, the Financial Times reported that Legora, the Stockholm company that is Harvey's closest rival, had started early discussions above $10B, roughly double the $5.6B it reached in the spring, with one investor suggesting the figure could reach $12B. Neither round has closed at the reported price, and both companies declined to comment.

A price nobody has signed still records what the people at the table believe a business is worth, and these two conversations put roughly $25.5B on companies founded in 2022 and 2023. What the money is buying is revenue, and the two produce it from opposite directions.

Harvey is the larger business and the more conservative bet, with annualized revenue above $350M, which prices it at roughly 44 times its current annual earnings. Legora is the smaller business moving faster, having reached $100M in recurring revenue in April at a pace Bessemer Venture Partners called the quickest in the history of enterprise software, and reporting that, based on current talks, it puts it near $150M. That works out to between 67 and 80 times revenue, so the market is paying a steeper price for the faster curve than for the larger company.

What the price assumes

Multiples like those only make sense if Harvey and Legora are still sitting between the models and the lawyers many years from now, which requires the companies building the models to stay on their own side of the line. Through 2026, they have repeatedly walked across it, beginning in late January when Anthropic released a set of plugins for its Claude Cowork agent, one of which handles contract review and compliance work. CNN reported that Thomson Reuters fell 15.83% in a single session, the largest one-day drop in the company's history, as investors decided in one afternoon that a general-purpose tool could do work the legal software companies had been charging for.

The second crossing was larger, and in May, Anthropic launched a dozen practice-area tools along with connectors into Westlaw, DocuSign, and Harvey itself. Fortune reported that Freshfields, Quinn Emanuel, and Holland & Knight were already running Claude on live matters, which put lawyers at three of the largest firms in the world doing billable work inside a tool sold by one of Harvey's own suppliers.

The benchmark that proved the wrong thing

Selling into the same buildings is the half a competitor can see coming, while the harder half is that the supplier decides how much work is left over for anybody else to do. General AI benchmarks test reasoning puzzles and coding problems, which reveal very little about whether a contract summary is usable in front of a client, so Harvey built an internal test of real legal work to prove its founding claim: a general model tuned for law, by people who understand law, will beat the general model on its own.

The test ended up proving something closer to the opposite, because when Harvey explained why it had begun offering Anthropic and Google models alongside OpenAI's, the company noted that seven newer models had outperformed its tuned version in a single year. Nobody took that advantage away, and nothing about Harvey's own work got worse, since ordinary progress at the model companies caught up with it and kept going, which means the most valuable part of the product improves and expires on a schedule Harvey does not set. A company telling its customers which suppliers' models it now runs is doing what the assemblers did with the sticker.

None of which has slowed either business down, and the growth is the strongest evidence that they are building something the models cannot easily absorb, since Harvey added more than $100M of recurring revenue in a single quarter. Deep integrations, security clearances, and customer relationships that are slow and expensive to unwind are real assets, and Anthropic building a connector into Harvey suggests the model companies still treat these applications as partners rather than targets.

The buyers have started running their own test

Neither vendor can settle that argument, since both have a financial reason to believe their own version of the future, so the better signal comes from the law firms writing the checks. In May, Kirkland & Ellis, the highest-grossing law firm in the world, said it would spend $500M over three to four years building its own platform rather than buying one, designed so the underlying model can be swapped out without rebuilding everything around it. Kirkland is paying half a billion dollars for the right not to depend on any single supplier.

Very few firms can afford that, which leaves everyone else with a more basic question: how does a firm know whether the software it is buying is worth what it costs? Thomson Reuters, whose stock took that beating in January, found in its own survey that only 18% of professionals said their organization tracked the return on its AI spending, while another 40% did not know whether anyone measured it at all. Budgets approved in a hurry get examined carefully later, and the question then will not be whether AI is useful, but whether each vendor's software is useful enough, next to what the model already does on its own, to keep paying for separately.

What the middle layer actually owns

Harvey and Legora have answered that risk differently, and every company selling a specialized product built on models it does not own is choosing between the same two answers. Harvey has started building its own legal models, which is the choice Apple made with its processors: stop renting the part that decides how good the product can be. Legora has remained model-agnostic, retaining the ability to switch suppliers if one becomes a competitor. Neither answer removes the risk, because a company building its own models is competing with laboratories that are far better funded, while a company that can switch suppliers still depends on the ones it switches between.

Both answers leave that $25.5B doing something narrower than the number suggests, since neither round has closed and the figure is what a small group of investors has said out loud rather than money that has changed hands. What they have said is that the arrangement has years left in it, which is a different claim from saying it is safe, because paying 44 to 80 times revenue is a bet on duration that pays out if the current division of work lasts long enough for the revenue to compound.

The available facts establish a changed set of incentives rather than a decision by anyone to displace anyone, and what is being priced is not a capability either company owns. Harvey's own test showed its advantage expiring when nobody attacked it, and Legora's answer is to keep its options open rather than build something the model companies could not. What carries these valuations is the willingness of OpenAI, Anthropic, and Google to keep leaving this work to somebody else, which is a judgment about where those companies point their attention rather than about what anybody has built.

The companies that assembled personal computers were never pushed out of the business, and most of them still ship more machines today than they did 20 years ago. What happened to them was slower and far harder to see coming, because the money inside a laptop kept drifting toward the part nobody could substitute, and the firms that built everything else ended up competing on price. They kept putting the sticker on the box right through it, since the arrangement was never publicly broken and no launch event marked the day the value started moving. That is the part that makes a shift like this so difficult to price while it is still happening.

Weekend To-Do

• Steal Granola’s ‘game of inches’ playbook: Read how the AI notepad reached a $1.5B valuation not through one killer feature, but by obsessing over dozens of small product decisions.

• Read Harvey's model-mix note: Skim why Harvey started offering Anthropic and Google models, and what its internal legal benchmark actually found.

• Run a 20-minute substitution test: Take one task your AI vendor handles and try it raw in a frontier model; the gap is what you're really paying for.

Friday Poll

Headlines You Actually Need

• OpenAI loses its data center chief: Chris Malone left last week, pushing OpenAI's 2026 executive departure count to 13 as its IPO slips to 2027.

• Instagram will cut your Reel for you: First Draft trims clips and strips out pauses to assemble a rough edit in under 10 seconds, iPhone first.

• Taiwan indicts nine over AI server smuggling: Prosecutors charged NVIDIA and Supermicro staff over servers routed to China through Indonesia and Japan, with 56 more seized at the border.

Meme Of The Day

The Toolkit

• Together AI: Cloud platform for running and fine-tuning open-source models at scale, no GPU babysitting required.

• Descript: AI audio and video editor that lets you cut recordings by editing the transcript.

• Lavender: AI sales email coach that scores drafts live and suggests rewrites to lift reply rates.

Rate This Edition