Four new cards, two operators. Tom Tunguz published three separate posts between September 3 and 8, each approaching a different layer of the AI economy. Simon Willison read the same OpenAI data that prompted Tunguz and pulled a different thread. The cards cluster into two themes.
Two readings of the same OpenAI data
OpenAI released internal data on how its research team uses coding agents. Both Tunguz and Willison wrote about it. They noticed different things.
Willison's Model capability crossing a threshold triggers agentic adoption sharply, not gradually focuses on the shape of the adoption curve. Researcher spending on coding agents sat near zero through February 2026, climbed gradually to about $150 per day by June, then spiked steeply to $600 by late August. Six weeks of acceleration after months of near-flat. Willison's read: the inflection tracks internal access to the model later released as GPT-6 Astra. Teams did not gradually learn to use agents. They adopted them fast once the model was capable enough to be useful.
"my best guess is that's when internal employees gained access to the model later released as GPT-6 Astra"
The principle is threshold-gated adoption. Tooling can be available and awareness can be high, but if the model quality falls below the point where agents reliably complete tasks, latent potential stays latent. The implication for teams planning AI rollouts: expect a slow plateau followed by a fast jump, not a smooth ramp.
Tunguz's Reported 3x AI productivity gains are agent shift-work economics, not intelligence multiplication reads the same chart differently. He questions whether the 3x gain the data implies is a productivity story or an accounting artifact. Researchers logged 3.14 agent-workdays per 8-hour human shift because agents run overnight. Inference costs surged 40-fold. Top researchers burned $7,000 per day. And over half of 4-to-8-hour tasks still required human intervention. Real delivered output is closer to 2x after defect rates. The frame Tunguz uses: paying for a second and third shift of machine runtime, not tripling researcher intelligence.
The two cards are complements. Willison explains when adoption tips. Tunguz examines what the output number actually means after it does.
- Simon Willison, Model capability crossing a threshold triggers agentic adoption sharply, not gradually. Researcher spending on coding agents stayed flat for six months, then jumped 4x in six weeks when OpenAI employees accessed a more capable model internally.
- Tom Tunguz, Reported 3x AI productivity gains are agent shift-work economics, not intelligence multiplication. The 3x agent productivity gain at AI labs reflects shift-work economics: agents run 24/7 while humans work 8 hours, with real delivered output closer to 2x after high defect rates.
AI economics: pricing as data acquisition and infrastructure as credit risk
Tunguz's other two posts this week move from productivity measurement to economic structure.
A two-tier AI API that discounts inference for training rights replicates the ads supply chain model analyzes Meta's two-tier API pricing. The subsidized tier costs 92% less than the full-rate tier. The difference is training rights. Meta acquires customer prompts as post-training material, paying inference costs in exchange for labeled reasoning traces. Tunguz names the structural parallel directly: this is the digital advertising supply chain applied to AI. Search and social platforms vertically integrated behavioral data by offering free or subsidized access in exchange for behavioral signals. Meta is applying the same logic to inference.
"the difference in price tells us the value of the data"
For founders building on top of AI APIs, this pricing structure signals something about the underlying economics. The platform capturing training data at scale is funding model improvement while simultaneously subsidizing customer acquisition. Privacy-first providers who do not take training rights will price at the upper end of the market. The two-tier structure becomes a pricing signal about which providers are in a data acquisition phase.
AI infrastructure is financed with $4 trillion in projected debt, requiring 55% annual AI revenue growth to service it zooms out further. AI data centers are financed with 70% debt and 30% equity, following real estate project finance conventions. At the scale being planned, the total debt load reaches roughly $4 trillion, which Tunguz computes as 286% of US commercial paper outstanding, 143% of global private credit markets, and a 34% expansion of the US corporate bond market. Servicing this debt requires AI revenue to grow at 55% compounded annually, from $150 billion today to $1.35 trillion by 2030. Tunguz's framing: this is no longer a corporate earnings story. It is a macroeconomic credit event.
"Financing the AI infrastructure boom is no longer a venture capital or corporate earnings story. It is a macroeconomic credit event that will rival the largest debt expansions in financial history."
The card is relevant for any founder building on AI infrastructure who wants to understand the macro dependency underlying cloud AI pricing. If the revenue growth rate slows, infrastructure expansion contracts. The systemic nature of the debt load makes AI adoption a structural input into credit conditions.
- Tom Tunguz, A two-tier AI API that discounts inference for training rights replicates the ads supply chain model. Meta's two-tier pricing offers a 92% inference discount in exchange for training data rights, replicating the ad-supply-chain model that search and social platforms used to vertically integrate data acquisition.
- Tom Tunguz, AI infrastructure is financed with $4 trillion in projected debt, requiring 55% annual AI revenue growth to service it. $4 trillion in projected AI infrastructure debt requires 55% CAGR in AI revenue to service; if growth slows, infrastructure expansion contracts and cloud AI pricing tightens.