The Infinite Compute Mirage

In June, legal tech darling Harvey saw its gross margins collapse from 50% to negative 50%. The startup was burning through compute so fast that every time a lawyer asked its agentic workflows to analyze a matter, Harvey effectively paid for the privilege of serving them. The vendor eventually patched the bleeding by swapping underlying architectures to an open-weight model post-trained on Moonshot AI's Kimi K3, but the underlying arithmetic remains exposed.

We are watching a massive, industry-wide shell game. The narrative sold to managing partners was simple: replace expensive human hours with high-margin software subscriptions, pocket the difference, and watch efficiency soar. Instead, agentic AI tools fan out across firm repositories, consume trillions of tokens, and generate astronomical cloud infrastructure bills. The overhead hasn't disappeared. It merely shifted from associate salaries to server farms.

The tech industry didn't liberate legal practice from tedious labor. It simply replaced associate overtime with a permanent tax paid directly to cloud providers.

The Illusion of Software-Scale Margins

The bull case for legal tech rests on the assumption that vendors can achieve traditional enterprise software economics. Legora is reportedly seeking funding at an $8.5 billion valuation after passing $100 million in ARR, while Chamelio secured $26 million to replace contract management systems with autonomous agents. Meanwhile, reporting by CodeHistory highlights how Harvey reached a $15.5 billion valuation backed by 80% of the Am Law 100.

That market exuberance ignores basic unit economics. Traditional SaaS platforms scale with near-zero marginal cost per user. Agentic legal tools do the exact opposite. As TBPN Digest reported, Harvey’s token consumption spiked 20-fold as lawyers transitioned from basic single-prompt queries to reasoning-powered agents that churn through entire matter histories. Selling flat per-seat subscriptions on top of hyper-metered compute foundations is a guaranteed way to lose money.

Firms Are Funding the Compute Burn

Law firms are not insulated from this burn; they are aggressively bankrolling it. Morgan & Morgan committed $1 billion over a decade toward its proprietary MX2 platform, while Kirkland & Ellis pledged $500 million to build internal tools alongside Palantir. These are eye-watering figures spent to build internal engines that face the exact same token-consumption realities as commercial vendors.

As bestlawfirms.com noted, recent industry commentary highlights that legal AI adoption has done nothing to kill the billable hour, with law firm revenues climbing nearly 12% in early 2026. The real cost burden sits in implementation and compute. When an agent loops endlessly through discovery sets, someone pays for those tokens. Today it is the venture-backed vendor taking the loss; tomorrow, those costs will be passed down directly to law firm tech budgets via metered pricing and hard usage caps.

A Cloud-Provider Tax on the Practice of Law

Worse still, these runaway compute costs buy zero immunity from professional liability. The Tenth Circuit recently proposed a rule requiring lawyers to certify human verification of AI-assisted filings, spurred by hallucinated submissions that earned one attorney a $1,000 sanction. Federal courts are making clear under Rule 11 that no amount of agentic token burn excuses a lawyer from independently verifying the record.

So where does that leave the working practitioner? You are paying thousands per seat for software that burns through compute like a furnace, only to spend your own unbillable hours double-checking its math to avoid getting sanctioned by a federal judge. The tech industry didn't liberate legal practice from tedious labor. It simply replaced associate overtime with a permanent tax paid directly to cloud providers.