This week's material spans compute investment, Tesla's earnings, and application-layer business models. Indigo pieces them into a single picture: models are becoming a raw material anyone can get. What is truly scarce sits below the model — power, chips, and data — and above it, the layer of organization that delivers results to the customer. This is a letter about foundations.
2026.04.19 — 2026.04.26 · Once a week: spot the signals, recalibrate.
This Week's Signals
Start with the facts. Google confirmed another $40 billion investment in Anthropic, along with 5GW (gigawatts, a unit measuring the scale of power and compute) of compute resources. Tesla's 2026 Q1 earnings showed that its next-generation AI chip, AI5, taped out early (sent to the fab for trial production after the design was finalized), with priority going to the Optimus humanoid robot and data centers rather than cars; the structure of TERAFAB, its plan to build its own chip fab, was confirmed, with full-year capital spending above $25 billion. On the enterprise side, nearly two thirds of companies are experimenting with Agents (AI programs that can call tools on their own and complete tasks for people), but fewer than 10% can scale them into real value. The SpaceX camp also acquired the AI coding tool Cursor for $6.7 billion.
Taken alone, these look like unrelated news items. Taken together, Indigo reads one logic in them: model capability is no longer the constraint. The constraint has moved to the two ends of the model — below it, the heavy-asset foundations of power, chips, and data; above it, the software layer that wires models into enterprise workflows (known in the industry as the harness) and the orchestration layer (the scheduling logic that routes tasks to the right models and tools). In McKinsey's survey, 80% of companies named data problems as their top obstacle. Anthropic's postmortem on its own Claude Code found the same thing from the other side: with the model unchanged, just three small issues stacking up — degraded inference, a cache bug, and prompt handling — made users feel it had gotten dumber. The quality of the experience depends on the architecture around the model, not the model itself, and this is where full-stack players are pulling away from single-point companies.
AI's next race is not a model race. It is a foundations race — whoever holds power, chips, and data at the same time defines the landscape after 2026.
Trends
#01 Two thirds experimenting, fewer than 10% delivering: the gap is the data foundation
Inside the enterprise Agent boom there is a glaring gap. Indigo said on X: "Nearly two thirds of companies worldwide are already experimenting with Agents, but fewer than 10% can truly scale them and produce real value. The problems almost all point to the same place — the data foundation isn't solid" (original post). McKinsey's survey backs this up: 80% of companies name data problems as their top obstacle, and the report lays out a five-layer data architecture and seven principles as the remediation path — a checklist that could almost serve directly as a due-diligence questionnaire for application-layer companies. Anthropic's self-review provides evidence from the other side: the model did not change, the engineering details around it did, and users felt it got dumber. The real bottleneck is not that models are not smart enough. It is that enterprise data cannot catch what the models throw. The 90% gap between PoC (a small proof-of-concept pilot) and production will be the densest zone of application-layer value over the next 12-24 months — the data and orchestration layers, plus the enterprise harness, are the next structural beneficiaries after compute.
#02 $40 billion and 5GW: harvest time for full-stack integrators
The loudest signal of the week came from Google. Indigo posted: Google has officially confirmed another $40 billion investment in Anthropic… and will also provide 5GW of compute resources
(original post) — his most widely shared post of the week. Google also holds a 6% stake in SpaceX. The same week, Tesla's earnings confirmed TERAFAB — Tesla putting up $3 billion for a research fab, SpaceX leading production, on Intel's 14A process — with full-year capex (spending on plants and equipment) above $25 billion and negative free cash flow. Nuclear company X-Energy secured a 5GW long-term offtake agreement with AWS running through 2039 (a contract locking in future generating capacity in advance), with AWS leading its $500 million Series C-1 round and the US Department of Energy's ARDP program adding roughly $1.23 billion in matching funds over seven years. Meta has already deployed tens of millions of Amazon's in-house Graviton CPUs. Google, Tesla, SpaceX, and Amazon are, almost simultaneously, pressing resources into every layer at once: power, chips, data, models. Indigo's core judgment this week: in Q2 2026, the full-stack integrators are pulling away from the single-point companies.
#03 Sell outcomes, not tools: the common formula behind FDE and Autopilot
This week Indigo turned application-layer business models into a single formula. In a long post (original post) he carried Palantir's FDE model (Forward Deployed Engineer — stationing engineers at the customer's site and reshaping the product around the customer's business) into the AI era. ColdIQ reached 7M+ ARR (annual recurring revenue) in 31 months without ever raising money. OpenEvidence went from 0 to a $12 billion valuation in 18 months, with 40%+ of US physicians using it daily, 20 million queries a month, embedded directly into the electronic medical record system of the healthcare provider Sutter Health. What these cases share is Autopilot rather than Copilot — the AI delivers the result and the human only signs off, instead of the AI assisting a human operator. As Jakob Nielsen put it: users go from operators to supervisors. Data moats come before software moats, and Indigo is blunt about services that have only logic and no data: "Booking, Tripadvisor… are all just databases of human activity. If your service can't even offer a database — if all you have is a logic layer — sooner or later an Agent will replace you" (original post). These companies sell outcomes, not tools. Following this thread, he and a friend also sketched out the early shape of an OPC (one-person company) fund — a vehicle that pushes the FDE idea to its limit.
On the Ground
#04 AI5 no longer goes to cars first
Indigo posted: "One chart to read Tesla's 2026 Q1 earnings… AI5 taped out early, mainly for Optimus + data centers (no longer cars first); TERAFAB structure confirmed" (original post). Tesla's scarcest chip goes first not to cars, but to robots and inference compute. Where the chips flow is where the company's center of gravity is.
#05 Upgrading rockets like software
After watching the Starship third-anniversary documentary, Indigo wrote: "The official 25-minute documentary for Starship's third anniversary… SpaceX can upgrade rockets as fast as software" (original post). Iteration speed is itself a moat. An organization that can change a rocket like shipping a release can apply the same cadence to chips — a hidden dividend of full-stack integration.
#06 The subject of consumption is changing
Indigo restated an old view: Invest away from carbon-based consumption — only invest in silicon-based consumption! I think I shared this view in a livestream back in 2023
(original post). What is changing is not the volume of consumption but its subject. When Agents search, compare, and order on people's behalf, the companies supplying machine demand stand on the growth side.
#07 The white-collar timetable
Indigo reposted Dario's prediction (original post): 50% of white-collar jobs — entry-level lawyers, consultants, finance roles — will disappear within 1 to 5 years, and he also proposed an extra tax on AI companies. What Autopilot eats is not the software budget. It is the service sector's payroll. When AI delivers the result directly, the people replaced are the ones who used to deliver it.
#08 The Agent's wallet
Cloudflare released NET Dollar, a stablecoin (an on-chain currency pegged to the US dollar) for Agent networks. Adjusted stablecoin transaction volume reached $10.9 trillion in 2025, against Visa's full-year payment volume of $14.2 trillion — but real-world payments accounted for only $400 billion of it. After compute and data, payments are the third piece of the Agent economy. Only with all three does machine-to-machine commerce close the loop.
#09 The brain is within computable range too
Isaak Freeman left MIT to work full time on digital humans, and Indigo wrote a long quote-post on whether it is feasible: the human brain's compute requirement of roughly 6×10²⁰ FLOP/s (floating-point operations per second) already falls within reach of today's AI clusters, and a Chinese team once ran a coarse-grained simulation of 86 billion neurons on 14,012 GPUs. Brain emulation is no longer a philosophical question. It is an engineering schedule. The foundations logic holds on the life-sciences side too.
Slow Thinking
Indigo publicly changed his line on one thing this week: Apple. His view had been neutral — excellent operations, weak innovation, no need for harsh words. On April 20 he posted a long piece on X and made it final: Cook is a master operator, not a product visionary; Apple's soul is missing a spark; the broad consensus is that new blood is needed. The trigger was direct: the news that Cook will step down on September 1, on top of his long-accumulated observations of Apple's stalled innovation. From ambivalence to verdict — the clearest position shift of the week. One other change is not a change of mind but worth noting: FDE had only been a topic he raised repeatedly in person; this week it became a systematic public essay for the first time.
On the other side, the strongest counterargument to this week's theme: the 90% deployment gap may not be a structural barrier, just a time lag. Every general-purpose technology lags between pilot and production. If model capability keeps jumping, the next generation might digest dirty data directly and bypass data remediation, making the investment in rebuilding data architecture unnecessary. Likewise, full-stack integration means Tesla spending over $25 billion a year in capex with negative free cash flow — if AI demand grows slower than expected, heavy assets turn from moat into burden. The test is also clear: if enterprise Agent deployment rates climb quickly over the next few quarters without large-scale data remediation, or if asset-light pure-software companies start outgrowing the integrators again, this theme should be thrown out.
Indigo on X
"Nearly two thirds of companies worldwide are already experimenting with Agents, but fewer than 10% can truly scale them and produce real value. The problems almost all point to the same place — the data foundation isn't solid"
From @indigox, 225 likes
"One chart to read Tesla's 2026 Q1 earnings… AI5 taped out early, mainly for Optimus + data centers (no longer cars first); TERAFAB structure confirmed"
From @indigox, 76 likes
"The official 25-minute documentary for Starship's third anniversary… SpaceX can upgrade rockets as fast as software"
From @indigox, 52 likes
Closing
One thought
At the end of the nineteenth century, electric motors began replacing steam engines, yet factory productivity barely moved for decades afterward. The gains arrived only when factories redesigned their entire production lines around electricity. The problem was never the motor itself — it was that factories were still laid out by steam-engine logic. Agents and enterprise data stand in the same relationship today: two thirds of companies have wired up the electricity, fewer than 10% have rearranged the shop floor. The model is the motor; data architecture and process are the shop floor. History keeps proving that the dividend goes to those willing to tear up the shop floor.
One thing to try
Spend 30 minutes giving one company a quick checkup. Pick a software or service company you use daily or have followed for a long time, and write down answers to three questions. One: does it own an exclusive database no one can take away, or does it only have a logic layer? Two: does it sell a tool (Copilot — the human operates) or a result (Autopilot — the machine delivers, the human signs off)? Three: if a model twice as strong appeared tomorrow, would its moat get thicker or thinner? If two of the three answers point to fragility, the divergence this letter describes is happening to that company right now.