AI the News That's Fit to Prompt

Two hundred and fifty rupees an hour, about two dollars, is what a worker in Tamil Nadu earns to strap a phone to their head and film themselves washing dishes ninety times a day so a robot can learn the motion. Six billion dollars is what Anthropic is reportedly willing to pay for a hundred-person Israeli startup whose main product is making the same chips go further. Those two numbers are the ends of the same rope. Today's issue is about what happens when the binding constraint on AI stops being intelligence and becomes supply: of compute, of electricity, of demonstration data, of credit, and of the human beings who generate all four.

The scarce thing is efficiency, and it is now worth billions

• Anthropic's reported $6 billion move on Decart is a bet on inference, not on a product. Decart was founded in 2023 by Dean Leitersdorf and Moshe Shalev, who met in Unit 8200, and employs roughly 100 people. Its core software makes AI chips run more efficiently, cutting the cost of training and serving models, and reports say the team would join Anthropic's inference and performance organization. It also builds world models: Lucy for real-time video transformation, Oasis for simulated environments used to train robots and autonomous vehicles. The company has raised about $450 million, including $300 million in May at a $4 billion valuation, up from $3.1 billion roughly eight months earlier. Nvidia is both a backer and a company that explored buying Decart outright before talks collapsed. Paying a 50 percent premium over a three-month-old mark, while preparing for an IPO, is a statement about how tight compute actually is.
• ASUS is the demand side of that same equation, measured in shipped racks. The company raised its full-year server revenue growth target for the second time in 2026, from roughly 50 percent at the start of the year, to at least 100 percent after Q1, now to at least 150 percent. Q2 consolidated revenue passed NT$200 billion for the first time, server revenue grew nearly 200 percent year over year, AI servers are about 90 percent of the server line, and Nvidia NVL72 rack-scale products are 30 to 35 percent of the mix. Co-CEO Hu Shu-pin credited "extremely high order visibility." When a hardware maker revises the same number up twice in eight months, the shortage is not a narrative.
• China's memory industry crossed two thresholds on the same day. CXMT is now worth roughly $524 billion against Tencent's $510 billion, making a Hefei DRAM maker the most valuable company in China less than three weeks after a STAR Market debut that raised $8.6 billion and closed up 466 percent. Separately, Counterpoint has YMTC at 14 percent of global NAND bits shipped in Q2, passing Kioxia and Micron for third place for the first time. The revenue picture is more sober: YMTC still ranks only fifth by revenue, because its bits go into consumer and low-priced OEM channels rather than the enterprise SSDs that now make up 48 percent of NAND volume, nearly double a year ago.

The physical bill is arriving, and somebody has to underwrite it

• India just doubled its own estimate of what this costs in electricity. The Ministry of Power told Parliament that AI data centers will add 26.3 GW of demand by fiscal 2031-32, against an earlier figure of 13.56 GW. AI-specific capacity is projected to grow roughly 24-fold, from 275 MW to 6,546 MW, with total operational capacity reaching about 12 GW by 2030 from roughly 1.5 GW in 2025. The ministry says renewables will cover it. Grid Controller of India chairman Samir Chandra Saxena has warned that the variable, spiky nature of data center load and its sharp ramps pose real stability risks, which is a different problem from generation and does not get solved by adding capacity.
• Private credit has become the AI build-out's lender of record, which is exactly where the exposure is hardest to see. AI-related private credit loans passed $200 billion in early 2026, with roughly $800 billion more of data center financing expected to route through private credit and off-balance-sheet structures over two years, against about $570 billion of total AI-related debt issuance projected this year. Origination is concentrated in a handful of managers, Blackstone, Blue Owl, Apollo, Pimco, BlackRock, lending through SPVs to neoclouds and developers, then securitizing outward until pension funds and 401(k) products hold the tranches.
• The second AI risk in that book has nothing to do with data centers. Morgan Stanley puts direct lending exposure to enterprise software near 26 percent, concentrated in precisely the sector investors expect agentic AI to disrupt. Headline default rates read about 2.1 percent, but adjusted for liability management exercises and shadow distress the effective rate is closer to 5.4 percent, and roughly $51 billion of software debt rated B- or lower matures in 2028, with another $50 billion in 2029.

The training data is people, and the people are cheap

• Humanoid robots need to watch humans work, so an industry has grown up to film them. In Los Angeles, the platform Sunain gives headbands with phone mounts to more than 1,400 contributors who record scripted household tasks. In Chennai and Karur district, workers wear phones strapped to their heads, GoPros, smart glasses, RGB and depth-sensor cameras, and motion bands on wrists, hands, and legs, for firms including Objectways and Bengaluru-based Humyn Labs. Pay is about 250 rupees an hour; one studio worker films roughly 90 videos a day at four minutes each. The output is ground truth for Large Behavior Models that map human motion to robot instructions.
• It reaches skilled work too. At the Lotte Hotel Seoul, staff wear body cameras while folding napkins, feeding a motion library built by the South Korean startup RLWRLD. Motion-capture suits, exoskeletons, and full-body teleoperation are described as the fastest-growing robot data category of 2026.
• Both sides said the quiet part out loud. Objectways chief executive Ravi Shankar: "Some jobs are supposed to be taken over, so humans can go and do better things." A worker named Sriramyachandra: "Who else will give you 250 rupees an hour just for doing housework?" Neither is wrong, and that is the difficulty.
• Thrive Holdings is the white-collar version, capitalized. It raised $2 billion at a $12 billion valuation from SoftBank, D1 Capital, and Altimeter, the first outside money into a vehicle Thrive Capital launched in 2025. OpenAI took a stake in December 2025 and embedded its own employees with portfolio companies. The platform spans more than 70 businesses: Current in accounting, with over 50 firms and 2,000 professionals, and Shield in IT services with about 20. Thrive says TaxAI has processed more than 7,000 returns at 98 percent accuracy and cut preparation time about 30 percent, with help desk resolution 36 times faster and the count of deployed agents doubling in a month. A third vertical will target physical assets: "data centers, manufacturing, healthcare, power, water, transportation," founding member Anuj Mehndiratta said.

The reckoning is procedural, not philosophical

• Twenty-nine state attorneys general went to trial against Meta on Wednesday in Oakland. The bipartisan coalition is before Judge Yvonne Gonzalez Rogers, led by Colorado, Kentucky, California, and New Jersey, alleging Facebook and Instagram were engineered to addict minors, that Meta misled the public about safety, and that it collected children's data in violation of COPPA. Opening statements are August 18, the trial should run about seven weeks, and both Mark Zuckerberg and Adam Mosseri are expected to testify. Meta has said damages could reach $1.4 trillion. It follows a $567 million judgment in New Mexico.
• Meanwhile Australia's under-16 ban shows what enforcement actually yields. Meta says it has removed about 756,000 accounts since the law took effect on December 10, 2025, comprising 462,000 on Instagram and 294,000 on Facebook, up from 504,000 in January. It is using AI analysis of profile signals like birthday posts and school-grade references, photo-based age estimation, and activity inference. And yet government figures and independent studies found more than eight in ten under-16s were still on social media in the ban's first three months. The eSafety regulator is weighing an enforcement suit; proposed maximum penalties would rise to A$99 million.
• Detection is spreading faster than agreement about it. Pangram raised $9 million led by Menlo Ventures in late July and shipped Pangram 4 plus an image detector, claiming better than 99 percent accuracy and roughly a one-in-ten-thousand false positive rate, with Maryland and Chicago researchers rating it among the best available. NewsGuard integrated it in March and has used it to identify more than 3,000 AI content farms. On July 21 Substack shipped Pangram-powered detection across posts, notes, and comments over 100 words. Writers called it a witch hunt within the day.
• And the state is now using the same class of tooling on trade. A White House trade report released today names more than 40 countries in China-linked transshipment routing, including Mexico, Canada, the EU, India, Japan, and South Korea. Building on Executive Order 14411 from June, CBP's Automated Targeting System runs machine learning trained on historical seizures and worldwide trade data over routing histories, claimed production capacity, ownership relationships, and port X-ray imaging. Exiger estimates $75 billion of goods were illegally transshipped in the year to February 2026, costing $19 billion to $34 billion in duties.

The Throughline

The connective tissue in today's issue is that every remaining bottleneck in AI is a supply problem in the physical world, and each one is being solved by buying it from someone who has less leverage than the buyer. Anthropic cannot manufacture more silicon, so it bids $6 billion for a company that makes existing silicon go further. ASUS cannot conjure racks, so it revises guidance and lets order visibility do the talking. India cannot summon 26.3 GW, so it defers to renewables and hopes the ramp rates behave. Robotics companies cannot synthesize the motion of a human hand scrubbing a toilet, so they pay two dollars an hour for it. These are not analogous problems dressed up in a theme. They are the same problem at four different points in a single supply chain, and the price of each input is set by how badly the buyer needs it and how few alternatives the seller has.

What makes the Decart story the right lead is that it prices the constraint precisely. Decart's valuation went from $3.1 billion to $4 billion in about eight months, and Anthropic is reportedly willing to pay $6 billion now, for roughly a hundred people whose product is efficiency. That is not a bet on a consumer surface or a frontier capability. It is a company deciding that the marginal dollar is better spent on getting more inference out of the chips it already has than on almost anything else. Read against ASUS raising its server growth target twice in eight months, the two stories triangulate the same fact from opposite ends: hardware supply is growing at a pace that would be extraordinary in any other industry, and it is still not enough, which is why the software that stretches it commands a premium usually reserved for the hardware itself.

The financing layer is where the tension actually lives. Private credit is now originating the data center build-out because public bond markets pushed back on volume, and that migration has a specific consequence: the exposure moves from marked-to-market instruments held by people who trade them into appraisal-valued instruments held by people who cannot easily sell. Morgan Stanley's 26 percent software concentration is the sharpest version of the irony. The same funds lending against AI infrastructure are heavily lent into the enterprise software companies that AI is most credibly expected to disrupt. A 2.1 percent headline default rate that becomes 5.4 percent once you count liability management exercises is not a crisis signal, but it is a measurement problem, and measurement problems in credit tend to resolve all at once rather than gradually.

Set the Tamil Nadu footage against the Meta trial and the shape of the year gets clearer. In Oakland, 29 attorneys general are arguing that a company knowingly engineered a product to capture minors, and Meta's own defense filing floats $1.4 trillion in potential damages. That is a decade-old harm finally reaching a jury. Meanwhile, in the present tense, workers are being paid two dollars an hour to generate the training data for systems built to replace them, under no legal framework whatsoever, and the CEO of the firm collecting it will say on the record that some jobs are supposed to be taken over. The lesson of the Meta case is not that the law arrived. It is that it arrived roughly fifteen years late, after the harm was fully realized and the business was fully built.

The Bigger Picture

The AI industry has finished the transition it started two years ago: it is no longer a software business that needs a lot of computers. It is an industrial supply chain with a software company at the top of it, and the economics now behave accordingly. Industrial supply chains are governed by the scarcest input, and today's issue is a tour of what those inputs currently are. Efficiency software, priced at $6 billion. Rack-scale servers, growing 150 percent a year and still short. DRAM and NAND, now a Chinese national champion worth more than Tencent. Grid interconnection in a country that just doubled its own load forecast. Debt, routed through private credit because the public market said no. And human demonstration data, which is the only input on that list with a conscience.

That last item is the one worth sitting with, because it is structurally new. Every previous wave of automation trained on the exhaust of work already done: transaction records, digitized text, images already on the internet. Physical AI cannot do that. The motion of a hand folding a napkin was never recorded, so it has to be commissioned, which means it has to be paid for, which means somebody has to be hired specifically to demonstrate a task in order for a machine to stop needing them to do it. Sriramyachandra's question, who else will pay 250 rupees an hour for housework, is not naive. It is an accurate description of a labor market where the highest-paying available work is participating in your own substitution, and where declining to participate does not stop the substitution, it just forfeits the fee.

The regulatory picture argues that this will not be corrected in time. Australia passed the most aggressive youth social media law in the world, Meta deployed AI age estimation to enforce it, 756,000 accounts came down, and more than eight in ten under-16s stayed on anyway. That is a well-designed rule with real penalties and cooperative enforcement producing a partial result. The Meta trial in Oakland is the alternative path, liability after the fact, and it is arriving a full platform generation late. The one place government is moving quickly with AI is CBP's transshipment targeting, where the state is the buyer rather than the referee. Enforcement capacity is being built fastest where it collects revenue.

The optimistic read is available and worth stating. Thrive's numbers, if they hold, describe real productivity: 7,000 tax returns at 98 percent accuracy, preparation time down 30 percent, help desk resolution 36 times faster. That is the case for this whole enterprise, and it is not nothing. The question the rest of today's issue raises is who ends up holding each piece of it. The efficiency gain accrues to Thrive's investors and OpenAI. The compute premium accrues to Decart's shareholders. The memory windfall accrues to CXMT's, and to a Chinese industrial policy that has been remarkably patient. The grid load accrues to Indian ratepayers, the credit risk to pension funds who did not pick it, and the training data comes from people earning two dollars an hour. Every one of those transactions is voluntary. That is precisely why the outcome does not require anyone to have behaved badly.

What to Watch

• Whether the Decart talks survive contact with Anthropic's IPO timeline. Nvidia already walked away from acquiring Decart once, and Amazon and Nebius have been named as prior suitors. A $6 billion cash-or-stock deal for an inference-efficiency team, booked just before a public listing, is the kind of transaction that either signals extreme confidence in compute scarcity or gets restructured when bankers see it. Watch which.
• Whether the private credit measurement gap narrows or widens. The number to track is not the 2.1 percent headline default rate but the spread between it and the 5.4 percent effective rate. If that gap keeps growing while $51 billion of sub-B- software debt walks toward its 2028 maturity, the AI infrastructure exposure and the AI disruption exposure inside the same funds stop being separate stories.
• Whether the Meta trial produces a discovery record on design intent. Seven weeks with Zuckerberg and Mosseri on the stand will generate internal documents about how engagement mechanics were built for minors. Whatever the verdict, that record becomes the template every subsequent AI product-liability case borrows from, and there is nothing comparable on the books for systems trained on commissioned human demonstration.