$150 for a lawyer to review an NDA was never about the work. It was about the margin. A junior on a fixed salary. A few minutes of attention and a clean billable entry. Firms survived on that spread for decades. Now the same task costs fifteen cents, and everyone wants to believe the surplus returns to the client or fades into efficiency. It doesn’t. The $149.85 leaves the profession. It flows to chips, compute, and the infrastructure that now performs the labor. The value migrated.
This is not a pricing story. It is a balance sheet story. The work still exists. Legal no longer holds the margin. This is a map of that migration.
The Margin Shift
Everyone wants to talk about the $150 NDA becoming a fifteen-cent task. They want to debate whether AI can draft it “well enough” or whether lawyers will lose work or whether clients will finally get the efficiency they’ve been begging for.
None of that matters. The real story is the movement of cost and margin.
The old model was straightforward. A junior associate on a fixed salary meant the marginal cost of an NDA was effectively zero. Once the work hit the queue, the firm printed profit from the spread. It wasn’t efficient, but the inefficiency was structural. The whole system moved at the same speed. There was no real alternative.
AI destroyed the bottleneck. The task still exists. The risk still exists. The client still wants someone to sign off. But the bottleneck that extracted margin is gone. And once that bottleneck disappears, the economics have to settle somewhere new.
AI moves the cost from labor to infrastructure. Compute cycles. GPU time. Token windows. Routing. Monitoring. Orchestration. All the things lawyers don’t understand and don’t want to think about. But this is where the margin goes.
LegalOn’s 2025 NDA benchmarks show the same shift. Costs drop 70 to 85 percent. LawTech vendors are candid about why: the savings route to compute, not to the client.
The quick NDA review might feel cheap to the client, but the base layer where production happens is not. The marginal cost of one more inference call is negligible. The platform that routes it is not.
The fantasy was that AI would make legal work near free. It didn’t. But it made the price transparent. Legal labor had opaque marginal cost. Silicon does not.
Where the Margin Landed.
The $149.85 surplus did not stay inside the profession. It did not improve partner economics. It did not fund apprentices or training. It did not strengthen firm stability. It did not expand access to justice efforts. It escaped the domain entirely.
NVIDIA’s Data Center revenue hit $51.2 billion in Q3 2025, up 66 percent year over year. Driven by inference workloads across enterprise verticals (including legal). Its full-year revenue crossed $130 billion. Its Blackwell chips alone generated $11 billion in Q4.
That is where the surplus went. Not into legal balance sheets.
AmLaw firms are still profitable in 2025. Many are thriving. But the point is the shifting of the margin on existing work. Most firms did not capture the upside of AI efficiency. They pay for it. AI tools, licenses, copilots, integrations. Law firms are customers of the new factory, not its operators.
AI did not make legal work free. AI made legal work a compute problem.
Once that happens, the center of gravity shifts. The spread migrates to silicon because silicon now runs the factory.
Clients Don’t Capture It.
The LawTech media likes to pretend AI is going to save clients money. Lawyers even try to sell that story. But the data says the opposite.
ACC’s 2025 survey is blunt: 59 percent of GCs report no savings from AI adoption. What they do report is faster cycle times, more throughput, higher output per headcount. Bloomberg’s 2025 analysis lines up, 91 percent of legal leaders say AI efficiency is used to increase pace, not reduce spend.
This is exactly what anyone who has worked in-house already knows. Companies do not redeploy savings into reduced budgets. They redeploy savings into velocity: faster product cycles, faster GTM bets, faster experiments, faster everything. What clients want is speed. If they get the NDA faster, they spend the surplus on the next acceleration. If a contract gets reviewed in half the time, they move twice as fast on the deal. Clients are not buying cheaper legal work, they are buying pace. So what does that leave for the profession?
AI does not reduce the cost of legal work, it exposes it. And once the cost is exposed, the value flows to the place that actually performs the labor.
The only viable move is to build the layer above the infrastructure: decision frameworks, judgment routing, escalation logic, contract intelligence, governance pipelines, compliance engines. GC as an operating platform.
The interface between risk and infrastructure is the only zone where value stays inside law. Not the document, not the hours, just the operating system. This is the only defensible position left.
Closing Note. This Is a Map.
This is not a warning, it is a description of where the economics already flow.
The future belongs to those who stop fighting for the $150 and start capturing the layer where the $149.85 actually goes: not the document, not the task, not the hours. The operating system.