In partnership with
In Todayโs Issue:
๐ธ OpenAI spent $34B in 2025 to make $13B
โ๏ธ Anthropic and the White House stay split on Fable 5
๐ AI layoffs hit a safety net most workers never claim
๐ Cursorโs 25-year-old CEO bets on Musk over his suppliers
๐ค BYDโs secret humanoid robot
โจ And more AI goodnessโฆ
โก The Signal
OpenAIโs leaked 2025 books show the staggering price of frontier AI: $34 billion spent to earn $13 billion, with reported losses near $38 billion.
Most of that loss is non-cash accounting rather than money out the door, but the cash burn is still steep, and it surfaces just as OpenAI moves to go public. The same week, Claude Fable 5 topped Epochโs capability index, Anthropic stayed locked in its standoff with Washington, and new data showed AI-driven layoffs hitting a frayed safety net. The models keep getting cheaper to beat on benchmarks while the companies behind them keep getting more expensive to run. The race for a smarter model is mostly won; the race for one that pays for itself is only starting.
All the best,
Kim Isenberg
๐ Cursor's 25-Year-Old CEO Bets His Future on Musk
Cursor CEO Michael Truell is tying his AI coding startup's fate to Elon Musk, even though it still runs on Anthropic's models. The 25-year-old MIT dropout, now worth an estimated $1.3 billion, struck a deal that pairs Cursor with SpaceX's Colossus supercomputer and gives SpaceX the right to buy the company for $60 billion later this year. The catch: Cursor's product still leans on Anthropic's Claude, the very supplier it is trying to outgrow.
๐ tl;dr: Cursor wants to escape its dependence on AI's giants by strapping itself to Musk's rocket.
๐ธ AI Is Cutting Jobs Faster Than the Safety Net Can Catch Them
Nearly 120,000 tech workers have been laid off in 2026 as companies chase AI-driven efficiency, and most who lose a job never file for help. Roughly 75% of unemployed Americans do not claim unemployment benefits, often because they assume they are ineligible, according to Bureau of Labor Statistics data cited by Fortune. The system is creaky too: largely unchanged since the New Deal, with some states cutting benefits from 26 weeks to 12 and wage replacement now below 30%. As leaders like Sam Altman and Dario Amodei warn of AI-driven displacement, the safety net meant to cushion it looks badly out of date.
๐ tl;dr: AI is accelerating layoffs into a safety net most workers don't use and that has not been modernized in decades.
โ๏ธ Anthropic and the White House Are Still at War Over Fable 5
Anthropicโs leaders flew to Washington on Monday and left without a deal: the White House still will not lift the export controls it slapped on Claude Fable 5. At issue is whether the modelโs guardrails can be stripped to expose the more dangerous capabilities of Anthropicโs Mythos system. The administration believes they can; Anthropic and a chorus of outside security researchers call the fear overblown. Luta Securityโs Katie Moussouris said guardrails are โspeed bumpsโ that โonly serve to slow down the less skilled.โ Commerce has signaled it could let Fable 5 back online for consumers, but only once the jailbreak concerns are resolved, leaving the most capable model on the market in limbo.
๐ tl;dr: The first real fight over who controls a frontier AI model is still unresolved, and it is writing the rulebook for every lab that comes next.
Treat every AI benchmark headline as a starting point, not a verdict.
Why it helps: Todayโs Graph of the Day shows Claude Fable 5 โsetting a recordโ on Epochโs Capabilities Index, yet its error bars overlap with three other models. Knowing how to read that keeps you from buying hype that doesnโt survive a second look, whether youโre picking a model for work or just scanning the news.
Try this: Paste any benchmark claim into your AI of choice and ask: โAct as a skeptical analyst. What is the confidence interval here, what context is missing, and what three questions should I ask before I trust this result?โ
๐ฌ Watch This
DeepMindโs documentary The Thinking Game follows Demis Hassabis and his team over five years, from the labโs early obsession with games and intelligence to the AlphaFold breakthrough that changed computational biology.
The film is also a rare look at where modern AI, in many ways, began: not as a consumer product, but as a scientific project to understand intelligence itself. It is less about hype and more about the slow, difficult process behind real discovery: ambition, uncertainty, failure, iteration, and the belief that AI could become one of the most powerful tools in science.
โIf you put a key under the mat for the cops, a burglar can find it, tooโ
โ Tim Cook, CEO of Apple, on why building law-enforcement backdoors into encryption leaves everyone less secure. (TechRadar)
While Anthropic spars with the Commerce Department over Fable 5, it is fighting a second, older front with Washington. Since February, the company has reportedly refused to let its models power fully autonomous weapons or mass surveillance, a stance that prompted Defense Secretary Pete Hegseth to brand it a โsupply chain riskโ and President Trump to order federal agencies to drop its products. A federal judge initially called the Pentagonโs move โtroublingโ and blocked it, but an appeals court has since let the designation stand while the cases proceed. Read together, the two standoffs suggest a government newly willing to bring frontier labs to heel.
OpenAI Spent $34 Billion to Make $13 Billion
The Takeaway
๐ OpenAIโs 2025 revenue roughly tripled to about $13 billion, per figures first reported by Ed Zitron and verified by the Financial Times.
๐ But spending hit about $34 billion, with $19 billion on R&D alone, and the reported net loss neared $38.5 billion, nearly eight times 2024.
๐ Most of that loss is non-cash: stripping out stock comp, Microsoft credits and a one-time conversion charge leaves a cash loss closer to $8 billion.
๐ The figures surface as OpenAI confidentially files to go public, putting its path to profitability under a microscope.
OpenAI is growing fast and losing money even faster. According to 2025 financials first reported by Ed Zitron and verified by the Financial Times, revenue roughly tripled to about $13 billion, but the company spent about $34 billion to get there, with roughly $19 billion of that going to research and development alone. The reported net loss came in near $38.5 billion, nearly eight times the $5 billion it lost in 2024.
Sam Altman, OpenAI CEO (Getty Images)
That headline figure is scarier than the cash reality. A large slice of the loss is non-cash: stock-based compensation, billions in Microsoft cloud credits, and a one-time accounting charge tied to OpenAIโs conversion from a capped-profit nonprofit into a public benefit corporation. Strip those out and the underlying cash loss is closer to $8 billion, still steep for a company earning $13 billion, but a very different story from โ$38 billion in flames.โ
Altman on stage at an OpenAI event (Getty Images)
The timing is not a coincidence: OpenAI has confidentially filed for an IPO and is reportedly preparing to go public this year, which is dragging its books into daylight. Zitron is blunt, calling the losses โastronomical, and far higher than most believed,โ and saying he is โnot sure how this company finds a way toward any kind of sustainability or profitability.โ OpenAIโs own internal projections lean the other way, forecasting deep losses through 2028 before a turn to profit around 2030.
Why it matters: OpenAI sets the pace for the whole industry, so its ability to turn record spending into a durable business is the question the entire AI market, and a looming IPO, is now pricing in.
Sources:
๐ https://www.wheresyoured.at/exclusive-openai-financials
No theory. No slides. Just pipeline.
Most founders know their product. Few know how to get it in front of the right people. In this hands-on session, Clay + HubSpot for Startups walk you through ICP definition, prospect list enrichment, and AI-personalized outreach. You launch your first sequence before the session ends. June 18. 11am ET / 4pm GMT.
The chart: Epoch AIโs Capabilities Index (ECI), a single score that rolls many hard benchmarks into one number, now puts Claude Fable 5 on top at roughly 161, just ahead of GPT-5.5 Pro (~160), GPT-5.5 (~158), GPT-5.4 Pro (~157), Claude Opus 4.8 (~156) and GPT-5.4 (~156).
The lesson: The model Washington just pulled from the market is, by Epochโs measure, the most capable one publicly benchmarked. Capability and availability are now moving in opposite directions.
The caveat: The race is tighter than the ranking looks. The 90% confidence intervals (the error bars) overlap heavily, so Fable 5โs โrecordโ over GPT-5.5 Pro sits within the margin of noise, not a clear win.
๐ค BYDโs Secret Humanoid Robot Wants to Run Its Stores and Its Factories
โก Bottom line: BYD, Chinaโs largest EV maker, has quietly run a humanoid-robot program codenamed โYao-Shun-Yuโ since 2022.
๐ก Why it matters: It throws the worldโs biggest carmaker into the embodied-AI race against Tesla, XPeng and a wave of Chinese rivals.
๐ What it means: Humanoids are becoming a manufacturing game, and BYDโs in-house batteries, motors and chips give it an unusual head start.
The next workers in BYDโs showrooms and on its assembly lines may not be human. The worldโs biggest carmaker has quietly spent four years building a humanoid robot, codenamed โYao-Shun-Yuโ after three legendary Chinese sage-kings, and it wants the machines greeting customers on the sales floor and bolting together cars on the factory line. Executive vice president Li Ke confirmed the program, run since 2022 inside BYDโs 15th Business Unit. His pitch: a carmaker is unusually well equipped for robotics, because BYD already mass-produces the motors, batteries, electronics, chips and precision parts a humanoid needs, and it can throw a 4,000-plus engineer autonomous-driving team and a planned 100 billion RMB AI budget at the rest.
(Pandaily)
โChinaโs robots lack a brain, while US robots have strong brains but weak limbs,โ Li Ke said, framing BYDโs goal as doing both well. The plan is an open platform other developers can build on, with BYD as its own first customer. And BYD is far from alone in the rush to staff stores and factories with machines: Tesla already runs about 50 Optimus units in Shanghai, and XPeng plans to mass-produce its eerily lifelike IRON humanoid, bound for its own stores and plants, by the end of 2026.
XPeng unveils its lifelike IRON humanoids at AI Day 2025; rivals are racing BYD onto the same showroom and factory floors (XPeng)
One AI employee. Engineering, finance, growth, ops.
Last week Viktor opened 14 pull requests, closed two month-end books, drafted a board update, deployed three landing pages, and triaged 600 support tickets. From inside Slack and Microsoft Teams. 20,000+ teams now run this way.
