Scott Wu says Cognition will make 30 to 35 million software engineers ten times faster. The number is precise to the million. The evidence is precise to nothing. Of the six claims that circulated around his remarks this week, four were vision, one was background, and one was a fact. That ratio does not describe a data set. It describes a term sheet with a press release attached.
I have audited this ratio before. In 2017 I hand-scraped 15,000 Ethereum transactions to test Tether's reserve claims, and the ratio was worse — one verifiable figure per dozen assertions. The ledger remembers what the press forgets. So let us do what the coverage did not: separate the claim from the arithmetic.
Context first, because the arithmetic needs a floor.
Cognition builds Devin, marketed as an autonomous AI software engineer. It is not a model lab. It is an orchestration layer — a planner, a tool-use loop, a memory store, a self-correction cycle — sitting on top of frontier models it does not own. In 2025 it acquired Windsurf, the IDE, after Google paid roughly $2.4 billion to license Windsurf's technology and absorb its CEO Varun Mohan and core engineering team. What Cognition bought was the shell: brand, users, distribution. The engine had already left the building.
Pricing tells the same story. Devin launched at $500 per month, enterprise-only. Devin 2.0 dropped the entry point to about $20 per month plus Agent Compute Units billed near $2.25 each. That is a migration from seat subscription to usage economics, and it is not generosity. It is the cost structure talking. Agentic workloads are token-dense: every planning step, every shell command, every retry burns inference. A flat seat price cannot cover a variable cost curve that steep.
Enterprise procurement is another variable the coverage ignored. Code review, compliance, liability, and audit trails do not compress at the speed of a demo. Efficiency hides the friction points. I watched the same pattern in DeFi: a protocol's technical throughput and its real adoption were different numbers, and the gap was always integration cost — wallets, custody, legal wrappers. For autonomous agents the integration cost is higher, because the output is executable and the failure mode is a production incident, not a failed transaction.
Combine Cognition's roughly $70 million ARR with Windsurf's roughly $80 million, and you get a $150 million run-rate against a reported $10 billion valuation. That is about 65 times sales. Mature SaaS trades at five to fifteen. Yields are just risk with a prettier name — and this yield is priced on the assumption that autonomous agents become the default enterprise delivery channel. That assumption has not been tested at scale. It has been priced at scale.
Now the core question, and the one the coverage skipped entirely: where does the 10x actually come from?
Trace it, and it does not come from Cognition. Devin's score on SWE-bench, the benchmark of real GitHub issue resolution, moved from roughly 13.9 percent in March 2024 to above 50 percent once Claude 3.5 Sonnet was plugged in, and past 70 percent with later frontier models. Audit the flow, not just the figure. The flow here is model capability. The orchestration layer contributes margin at the edges — better tool selection, better retry logic — but the steep part of the curve belongs to Anthropic and OpenAI. Cognition is renting the improvement and branding it.
The empirical range for coding assistance is not 10x. GitHub's randomized controlled trial put single-task speedup near 55 percent. METR's 2025 study of experienced open-source developers working in mature repositories found AI tooling made them 19 percent slower. That is not a rounding error; it is a sign inversion. The honest band runs between a 20 and 55 percent gain on isolated tasks, with negative returns on complex, context-heavy work. Ten times is not a benchmark result. It is an order-of-magnitude marketing claim with no published benchmark definition behind it.
For anyone trading this narrative on-chain, that distinction has a price. AI-agent tokens rallied through 2024 and 2025 on precisely this promise — autonomous software delivery, headcount compression, compute demand. Floor prices are narratives; volume is truth. When I mapped CryptoPunks wallet clusters in 2021 and found a single cluster wash-trading the floor, the volume looked healthy right up until it didn't. The same discipline applies here. Ask what the actual inference spend is. Ask what the retention is. Ask what fraction of Devin sessions complete without human rescue. None of those numbers have been published. Silence in the blocks speaks volumes.
Then there is the second-order trade that crypto markets have already priced. Agentic coding is token-intensive, and token-intensive workloads are compute-intensive. Every autonomous session that plans, executes, retries, and verifies is a metered draw on inference capacity. That demand flows to the same GPU and cloud infrastructure that DePIN compute networks have been trying to monetize. The connection is not direct — Cognition does not run on decentralized GPU — but the pricing signal is. If agentic coding genuinely scales, inference scarcity reprices the whole compute stack, and every token that sells compute exposure inherits the narrative whether or not it deserves it. Trace the coins, not the claims: check whether a compute token's actual utilization rose alongside the AI-agent headlines, or whether only its price did.
Here is the contrarian angle, and it cuts against the skeptics too.
The reflex is to say 10x is a lie, therefore the company is a lie. That is lazy reasoning — the same correlation-equals-causation error I watched analysts make during the 2020 DeFi summer, when I ran 10,000 Uniswap V2 iterations and found that most "impermanent loss disasters" were just volatility being repriced, not protocol failure. The protocol worked. The model was wrong.
So with Cognition: the 10x claim is unverifiable, but the underlying market is real. Enterprise software delivery is genuinely automating. If the cost of producing software falls, demand for software may expand faster than the labor it displaces — the Jevons effect. The certainty is not fewer engineers. The certainty is a restructured labor market, with junior, repetitive coding roles absorbing the first hit and architectural judgment moving up the value chain. That shift is already visible in hiring data, and it will show in salary dispersion long before it shows in headcount totals.
Where the risk actually sits is upstream. Anthropic and OpenAI already ship their own coding agents. They own the model and therefore own the marginal cost. Cognition's orchestration value can be internalized by its suppliers at zero incremental expense. That is not competition. That is a squeeze. The layer with the least capital intensity also carries the least defensibility.
The 2022 Terra collapse taught me the shape of this. When I ran the liquidation-cascade model across three lending protocols, the exposure was not in the headline yield; it was in the dependency graph nobody had drawn. Cognition's dependency graph runs straight into two companies that could erase it with a product launch. The ARR is real. The moat is a slide.
One more thing the ledger demands. Wu's framing treats the 30 to 35 million engineers as a market to be accelerated. It is also a labor force to be repriced. If a meaningful share of routine coding compresses into agent sessions, the demand for junior hiring falls before the demand for senior judgment does, and the pipeline that produces senior engineers thins at the bottom. That is a slow variable, invisible in a quarterly ARR line, and it will not surface until the shortage is already structural.
What to watch, concretely, over the next quarter — not the next year, because narrative pricing decays fast:
Watch ACU economics: if Cognition publishes inference cost coverage per ACU, the gross margin question gets answered, and if it does not, that silence is itself an answer. Watch the product cadence from Anthropic and OpenAI on agentic coding — every release there reprices Cognition's orchestration premium. Watch for a published benchmark definition behind the 10x claim: code volume, feature delivery, cycle time, whatever it is. A number without a denominator is not a metric. It is a slogan.
I spent 2024 building a dashboard that tracked 500,000-plus ETF flow data points against spot volatility and found a 0.85 correlation with shrinking exchange reserves. That number mattered because it was reproducible. Anyone could rebuild it. That is the standard. Wu's 30 to 35 million engineers deserve the same courtesy: not disbelief, not belief — a denominator.
The ledger is still open. The question is whether anyone at Cognition intends to write in it.
