Go to Groundby Jaidev Karthickeyan
← All notes Enterprise · February 2026

QuarryIf intelligence is co-created across organisational boundaries, who captures the upside?

Every control layer takes a cut. AI platforms want theirs on the outcome.

App stores take 15–30% of each transaction and cloud marketplaces about 3%. The next ask from AI platforms may be 1–2%, taken on the value their models help create.

What happened

Imagine a drug company that uses an AI model and doubles the success rate of its pipeline. The model provider sends a bill for tokens, a few tens of millions. The extra drugs are worth billions. Sooner or later, someone at the provider is going to ask why they are charging for the fuel and not for the oil strike.

For two years the enterprise argument about AI has been a tidy legal one: who owns the output? If a model writes code, designs a molecule or drafts a strategy, the assumption has been that the company owns it, as it owns a spreadsheet or a CAD file. That treats AI as a very clever tool.

The sharper question, to my mind, is who owns the intelligence that shaped the result. Frontier models don't just retrieve; they synthesise, from patterns built in pretraining, alignment tuning and reinforcement learning. When a platform starts talking about value-share or outcome participation, the change is not merely commercial. It turns AI from a tool you rent into a stakeholder that wants a cut.

The arithmetic

Pharma pipeline
Programmes200
Success rate, without and with AI3% → 6%
Expected approvals6 → 12
Net present value per approved drug$1.5bn
Incremental value from six extra approvals$9bn
1% participation on that increment$90m
Enterprise revenue
Annual revenue from AI-influenced products$5bn
Share that depends on AI-assisted design40% ($2bn)
2% participation on that share, per year$40m
Present value over 10 years at 8%over $260m
Running a private deployment instead, per year$50m

Illustrative scenarios. For scale, Wong, Siah and Lo (2019) measured oncology's success rate from first trial to approval at 3.4%, close to the 3% baseline used here.

Going to ground

First order

Pricing follows value. If a model lifts a pipeline's success rate from 3% to 6%, a token bill in the tens of millions looks absurdly small next to billions in added value. From the platform's side of the table, asking for a share starts to look rational.

Second order

Nobody can say who did the thinking. A share of the upside needs attribution, and attribution inside a large model is murky: there is no single "idea neuron" behind a breakthrough. In a real deployment, retrieval, fine-tuning on the company's own data and orchestration all shape the result, so the value is spread across the stack. Intellectual property stops being only the final artefact and starts to include prompts, embeddings and inference traces. Privacy moves from data at rest to intelligence in motion: even with no training on customer data, knowing which industries run long contexts or multi-step reasoning gives a platform a view across sectors that no single customer has. And if a platform shares the upside of an AI-assisted trading strategy, does it share the liability when the strategy fails in front of a regulator?

Third order

Surplus flows to whoever controls the layer. General-purpose technologies have a habit of this: cloud providers made their money on compute, mobile operating systems on distribution. If AI platforms become cognitive infrastructure, they may collect at the outcome rather than the usage. That makes architecture a financial hedge: $50m a year of private infrastructure is cheap if it protects hundreds of millions in enterprise value. Incentives shift too. Platforms paid on success may chase the richest industries hardest, and the alignment layers that steer models become commercially consequential. Intellectual property law assumes an identifiable author, and by the time regulation catches up, the norms may already be locked into multi-year contracts. Boards will need to read AI contracts for long-term surplus, not just cost.

Ears up: what to watch

  • Value-share or outcome-based clauses turning up in enterprise AI contracts.
  • Who owns the exhaust: contract terms on telemetry, metadata and inference traces.
  • Whether liability travels with the upside: does a share of the win come with a share of the loss?
  • The price of the hedge: private deployment against API pricing, as it gets cheaper or dearer.

Bone of contention

The comparison in the chart flatters my argument, because it isn't like for like. App stores and marketplaces charge on transactions anyone can count. Participation would be charged on outcomes someone has to attribute, and attribution is exactly what large models make hard. The pharma and revenue figures are illustrative scenarios, not observed contracts.

J. & the researcher

Sources
  1. UK CMA, mobile ecosystems market study, Appendix H: Apple and Google commission ratesconfirmed
  2. AWS Marketplace, listing fees: 3% on SaaS public offersconfirmed
  3. Microsoft commercial marketplace documentation: 3% store service feereported
  4. Steam revenue share tiers (30%, 25% above $10m, 20% above $50m), as widely reportedreported
  5. Wong, Siah and Lo, "Estimation of clinical trial success rates and related parameters", Biostatistics (2019)confirmed
  6. Participation rates of 1% and 2%: scenarios from the original articlescenario

Data for the chart: take-rates.csv · Originally published on Medium.