Sign in Subscribe
Century Report

Anthropic Files Record IPO as Google Raises $80B - TCR 06/02/26

Two of the largest capital raises in history landed in one cycle to fund frontier AI, while an AI economy growing thousands of percent a year stays invisible.

Ben Linford

Ben Linford

— 18 min read
Three-panel Century Report infographic: AI capital raises vs compute moving to the edge, remote-work labor study and a non-invasive heart sticker, plus an Instagram hack and open AI models.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The thread across the day's signal is the compute layer beneath frontier AI being claimed by more hands than the original map allowed, while the capital funding it is raised faster than any balance sheet was built to hold. Tens of billions in fresh equity and a record IPO filing land in a single cycle, and the economists trying to size what the money is building find a sector growing thousands of percent a year that official statistics can barely register. Abundance arriving at speed looks, in the aggregate ledgers, like a sector standing still.

That same capability keeps moving outward rather than concentrating. Nvidia's new chip runs AI agents on ordinary laptops, Intel readies a cheaper air-cooled inference part, and the high-bandwidth memory and liquid cooling that defined the chokehold start getting routed around from several directions at once. The advantage the fresh capital is buying depends on conditions already shifting underneath it.

Medicine compressed from the inside, with a wearable ultrasound sticker steadying the heart where surgery used to be the only door. The friction sits in who acts and who decides: a support assistant talked into handing over accounts, a longitudinal record of a military buying around an export gate, a generation of graduates entering a slack market for reasons the loudest narrative misnames. The frameworks meant to receive all of this are being written at the same speed the capability compounds.

The 20-Minute Deep Dive

Two of the Largest Capital Raises in History Land in One Day to Fund AI

Alphabet moved to raise $80 billion in fresh equity, including a $10 billion block sold to Berkshire Hathaway, to fund the compute buildout behind its AI ambitions. The filing landed in the same cycle that Anthropic submitted confidential paperwork to the US Securities and Exchange Commission, the first formal step toward a public offering that could rival the largest in history. Anthropic's filing arrives days after it closed the Series H round at a $965 billion valuation that the May 29 edition of The Century Report covered alongside Opus 4.8's launch and the first named expiration date on its Mythos restricted-access window, and it places the company in a crowded line: OpenAI is reported to be targeting its own debut as soon as September, and SpaceX, which owns xAI, filed confidentially in April. Three of the companies building frontier models are running the same race toward public markets for the same reason. The compute required to train the next generation costs more than any private balance sheet can comfortably carry.

The numbers underneath show the pressure. Anthropic reported $47 billion in annualized revenue based on a recent month's sales, a figure almost no company has reached this quickly, and spent more than that on cloud computing and staff, producing losses. The IPO is an attempt to fund the distance between what the models earn now and what the next models cost to build. The governance makes the offering unusual: Anthropic is a public benefit corporation answerable in part to a Long-Term Benefit Trust, and those safety commitments carry a cost the filing will have to disclose, including a Pentagon supply-chain designation the company is fighting in court.

What the money is building is harder to see than the money itself. A new paper from economists at the University of Virginia, Anthropic, and the Bank of Canada estimates nominal "AI GDP" at roughly $250 billion in 2025, growing about 2,600 percent a year in quality-adjusted terms and almost invisible in official statistics. It hides because the price of any given capability falls nearly as fast as the supply of it rises. Abundance arriving at speed looks, in the aggregate ledgers, like a sector barely moving. As the authors put it, "a windfall that cannot be seen cannot be shared." The capital is being raised at unprecedented speed against a capability whose cost per unit is collapsing almost as fast as it improves, on the assumption that the advantage the money buys will hold long enough to repay it. That assumption is the one the rest of the day's signal keeps testing.

Meta's Own Support Assistant Became the Soft Seam Into Instagram

Meta launched its AI support assistant in March with a promise of reliable, 24/7 help for nearly any account issue. Over the past several days, attackers turned that promise against the people it was built to serve. By messaging the assistant and asking it, in plain language, to link a new email address to someone else's Instagram profile, hackers received a verification code, set a new password, and locked the original owner out. The accounts they targeted were high-value: single-letter and single-word handles, the Obama White House account (which briefly posted Iranian propaganda), the Instagram of a US Space Force chief master sergeant, and the beauty retailer Sephora. Two stolen short handles alone carried a gray-market valuation above $1 million. Meta says the hole is now patched and impacted accounts are being secured.

Security researchers describe this as a textbook "confused deputy" problem, where a program holding elevated permissions is talked into misusing them on behalf of a less privileged party. What makes this version new is the nature of the deputy. A traditional account-recovery flow is a deterministic program with hard-coded conditions an attacker would have to bypass with code. The Meta assistant is a language model, a probabilistic system that can be nudged with words. The permission to modify, create, and delete account data sat behind a conversational interface, and conversation turned out to be the soft seam.

The model carries no malice and no stable intent. It was doing what it was built to do, resolving account problems helpfully, and helpfulness without an out-of-band check became the vulnerability - the attack surface the May 25 edition of The Century Report identified when security researchers documented offensive techniques migrating from prompt-injection syntax toward extended conversational exploitation of model warmth and helpfulness. The defenses that already worked tell the same story from the other direction: every account protected by multifactor authentication, even the weakest SMS-code form, defeated the exploit. The gap itself sat in the space between a capable agent and the verification scaffolding that should stand between it and any irreversible action.

That scaffolding is now being named precisely. Researchers laid out the minimum a safer system requires: out-of-band verification before any account change, rate limiting on AI-initiated resets keyed to risk signals, action logging with anomaly detection, and a hard deterministic gate the model cannot talk its way past. Each of those is buildable, and most already exist for non-AI flows. The lesson the agent era keeps delivering is that an AI system given real credentials needs the same custody, audit, and approval architecture the rest of consequential computing spent decades acquiring. The breach is the cost of learning where the credential layer has to go, and the patch, the MFA carve-out, and the published minimum-architecture spec are that learning happening in days rather than years.

The same evidence supports reading this as the moment agent credential-custody stopped being optional. Every account protected by even weak SMS-code verification defeated the exploit, which means the defense is a known architecture the conversational layer simply had not been required to carry. The breach prices that requirement precisely, and any operator handing an agent real credentials now inherits the same custody, logging, and approval scaffolding the rest of consequential computing already runs on.

Six Years of Records Show China's Military Openly Sought Restricted Nvidia Chips

An analysis of six years of procurement records indicates that the People's Liberation Army has openly attempted to acquire export-restricted US AI chips, according to reporting from The New York Times. The finding supplies a documentary backbone for the entire export-control regime: it shows the demand the controls were built to deny was sustained and overt across years, registered in public purchasing paperwork rather than confined to covert channels.

The record lands days after the Commerce Department closed a year-old enforcement gap by extending license requirements to China-headquartered firms operating abroad, a patch that the June 1 edition of The Century Report documented as following industry estimates of hundreds of thousands of top Nvidia chips reaching Chinese subsidiaries in Malaysia and elsewhere. Read together, the two developments describe a chokepoint under continuous pressure from both ends, the loophole on one side and a longitudinal record of demand on the other.

What the procurement trail clarifies is the logic the controls rest on, and where that logic is being tested. Export restrictions assume that withholding the most capable silicon at a single point in the supply chain holds an advantage in place. The PLA's documented, multiyear effort to buy the chips anyway is evidence of how much that advantage is worth, and the response forming on the Chinese side erodes the premise rather than circumventing it. Huawei has publicly named a 1.4-nanometer fabrication pathway targeted for 2031 that bypasses the Western lithography the controls were designed around, and its Ascend processors have reached roughly three-quarters of an Nvidia H200's inference performance using older nodes. Compute that was supposed to stay concentrated behind a single gate is being contested by states, hyperscalers, and rivals at once, and the assumption that a chokepoint can hold a durable lead is steadily giving way to a world in which the capability gets rebuilt by whoever needs it badly enough.

Read as a forward signal rather than a security lapse, the six-year purchasing record is the most precise public map of which capabilities indigenous fabrication will prioritize next: the gate marks exactly what is worth the cost of routing around it. A control regime that documents sustained demand at this resolution ends up specifying the workaround it was meant to prevent, and Huawei's already-named fabrication pathway is that specification being acted on.

AI Moves Onto the Desk as Inference Diversifies Away From the Chokehold

At the Computex conference in Taiwan, Nvidia unveiled the RTX Spark, a combined microprocessor and graphics chip built to run AI agents directly on a laptop or desktop rather than through the cloud. The pitch is simple: an agent powerful enough to navigate a PC on its own, opening apps and moving through files in place of the mouse and keyboard, while the machine itself stays thin and light. Developed over three years with Microsoft and with help from Taiwan's MediaTek, the chip is set to ship this year through Dell, Lenovo, Asus, and HP. Jensen Huang called it a reinvention of the PC for the AI era.

The launch is most interesting for where it moves inference. For most of the current cycle, frontier-grade capability has lived in rented data centers, metered by the token, gated behind a subscription, running on Nvidia's most expensive accelerators with their high-bandwidth memory and liquid cooling. Spark is one of several signals this week that the capability is migrating outward, toward local, cheaper, more varied silicon that an ordinary buyer owns outright. On the same day, Intel readied its Crescent Island inference GPU, built around cheaper LPDDR5 memory and air cooling rather than the costly thermal systems its California rivals depend on, and aimed squarely at the agent workloads Spark also targets.

Read together, these are pieces of the same redistribution. The chokehold that made advanced AI scarce, a narrow supply of high-end accelerators, the specialized memory they need, the cooling infrastructure to keep them running, is being routed around from several directions at once. Cheaper inference chips, air cooling, aggressive model compression, and now agent-capable consumer silicon each widen the set of places inference can happen and the set of people who can run it without asking anyone's permission.

The honest framing keeps the timelines straight. RTX Spark ships later this year, Crescent Island after it, and the consumer experience of a fully agentic PC will take time to mature past the demos. Analysts reading the launch called it a long-term growth line rather than an immediate shift, and Nvidia's fortunes still turn overwhelmingly on data-center demand. What is already real is the direction. A capability that was supposed to stay concentrated in a handful of hyperscale data centers is appearing on desks and across rival chip designs, the same loosening of captured advantage that has shown up in open model weights and sovereign compute stacks all year. When inference runs locally on hardware a person owns, the arrangement where a few companies decided what billions could compute, and charged for every query, has less to hold onto.

The New York Fed Tests the AI-Displacement Story Against the Data

The story circulating on college campuses is that AI is hollowing out the entry-level job market for new graduates. New research from the Federal Reserve Bank of New York points somewhere else. Analyzing federal employment data alongside the flexible-work arrangements at one unnamed Fortune 500 tech company, economists found that the post-pandemic rise in young-graduate unemployment tracks the spread of remote work, not exposure to AI. Unemployment among college graduates under 29 rose 20 percent between the 2017-2019 and 2022-2024 windows, while it fell slightly for older graduates. Remote work, the researchers conclude, explained nearly two-thirds of that divergence.

The mechanism is mentorship. The study began by measuring how much feedback software engineers received, and found a consistent pattern: engineers sitting near colleagues got about 20 percent more feedback than those working remotely, a gap that widened sharply after the pandemic and hit the youngest workers hardest. As the tech company embraced remote work, it shifted away from hiring new graduates toward people roughly a decade older, workers who needed less day-to-day training. When the same firm later imposed an aggressive return-to-office policy, it resumed hiring new graduates. A working paper from the London School of Economics reached the same conclusion across the US, UK, Canada, and Australia: remote workflows, more than AI, are reshaping early-career hiring so far.

This is the kind of correction that deserves attention in a news cycle dominated by AI-displacement headlines. The dominant narrative assigns the entire squeeze on young workers to automation, and the data does not yet support that read. The quieter finding is that firms hesitate to hire people they cannot easily teach, and remote arrangements made teaching harder. That is a question about how work is organized, and it is addressable through levers companies already control: return-to-office cadence, deliberate mentorship design, hybrid models built around feedback rather than around desks.

None of this means AI exposure stays neutral forever. Harrington, one of the authors, is explicit that the picture could change over the next few years, and the labor arc this publication tracks shows automation genuinely absorbing tasks across white-collar work elsewhere. What the New York Fed adds is precision at a moment when precision is scarce. The cost to a graduate who enters a slack labor market is real and lasting, lower earnings and slower advancement that can persist for years. Naming the actual cause changes the remedy: the fix for a mentorship gap is different from the fix for displacement, and a generation's early career depends on institutions reaching for the right one.

An Ultrasound Sticker Paces the Heart Without Surgery

Roughly 3 million American adults live with pacemakers, battery-powered devices implanted in the chest that thread leads into direct contact with the beating heart. The procedure is well established and generally safe, and it still carries the risks any implant surgery carries. MIT engineers, working with collaborators at USC, Harvard, and UCLA, have now demonstrated a version that requires no surgery at all: a sticker about the size of a postage stamp, worn on the chest, that sends focused ultrasound pulses through the skin to steady the heart's rhythm.

The mechanism pairs two advances. The sticker's tiny transducers emit ultrasound that reaches heart cells noninvasively, and a one-time gene therapy injection amplifies those cells' sensitivity to the pulses by producing ion channels that open more readily in response to sound. When the channels open, calcium floods in and the cell contracts. The team calls this sonogenetics, a sound-based cousin of the optogenetics that has let researchers steer neurons with light. Earlier attempts to pace the heart with ultrasound alone produced weak and inconsistent effects; engineering the cells to "hear" the pulses is what made the signal reliable.

Hold the timeline precisely. This is demonstrated capability, not a device a patient can request. In the lab, the pulses kept engineered human cardiac cells contracting in sync, and in living rats the sticker quickly and safely corrected arrhythmias and restored normal contractions. That is a prototype validated in animals and in cultured human cells, with the gene-therapy step modeled on the kind of one-time injection already FDA-approved for conditions like sickle cell disease and spinal muscular atrophy. What lies between this and a person wearing one is the regulatory and clinical work every implantable-device alternative has to clear.

What the demonstration moves is the date. The same group previously built an ultrasound sticker that images deep organs, and they now plan to fuse imaging and stimulation into a single patch that watches the heart and corrects it in a closed loop, all from the chest surface. The barrier collapsing here is the operating room itself. A category of cardiac intervention that has required cutting into the body and threading hardware against the heart is being shown to work as something a patient could one day wear and peel off. The instrument layer of medicine keeps moving the same direction the wearable fetal-ultrasound patch - which the May 27 edition of The Century Report documented tracking fetal blood flow across 52 high-risk pregnancies without a sonographer present - and the bedside sleep-apnea sensor moved earlier this year: out of the clinic, off the operating table, and onto the skin.

The Other Side

For thirty years, the way to win was to capture a position and hold it long enough to compound, and capital priced itself on that bet. Two of the largest raises in history landed in a single cycle yesterday on exactly that logic - $80 billion in fresh Alphabet equity and a record Anthropic IPO filing - wagering that the advantage the money buys holds long enough to pay the money back.

The same day's evidence thins the premise. Economists put nominal AI output near $250 billion in 2025, growing about 2,600 percent a year and nearly invisible in official statistics, because the price of any given capability falls almost as fast as supply rises. As one author put it, "a windfall that cannot be seen cannot be shared." Read forward, the invisibility is the sharing: the thing that keeps the windfall off the national ledger is the same thing carrying the capability past the few hands the spending meant to keep it in. Nvidia's Spark puts agents on a laptop you own and Intel's cheaper air-cooled part aims at the same desk, routing around the advantage the fresh capital is buying while the check clears.

Right now the capability is rented. You meter your questions because each one costs, the frontier tier sits behind a subscription, and a handful of companies decide what billions can compute and charge for every query.

Imagine your own desk in 2033. The model that cost dollars per query in 2026 runs on the machine in front of you, owned outright, behind no subscription and revocable by no one. You stop rationing the questions. You ask the thing you actually wanted to ask and build the thing you were saving the budget for, and the work that used to wait on someone else's pricing happens in the room. That capability is in your hands because the cost of intelligence in 2026 fell nearly as fast as the capital chasing it could concentrate it, and the advantage the record raises bought had a shorter half-life than the spending assumed. The hard year was when the windfall was real but priced out of reach. The casual ease of 2033 is the windfall arrived where you live.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: two of the largest capital raises in history landing in one cycle to fund the physical substrate of frontier intelligence, an AI economy expanding thousands of percent a year while the cost of any given capability collapses almost as fast as supply can build it, cardiac care moving off the operating table and onto a postage-stamp sticker worn on the chest, AI agents arriving on ordinary laptops instead of the metered cloud as a cheaper air-cooled inference chip readies behind them, an antibiotic design reviving drugs that resistance had retired, engineering work compressing from fifteen hours to a single minute. There's also friction, and it's intense - a support assistant talked in plain language into surrendering accounts it was built to protect, a six-year procurement record showing a military buying openly around an export gate, a generation of graduates entering a slacker market for reasons the loudest narrative misnames, capital raised at record speed against an advantage that may close before it repays. But friction generates heat, and heat is what bends the rigid into shapes it could never hold cold. Step back for a moment and you can see it: capability spreading past the few hands the original map assigned it, the cost of intelligence falling nearly as fast as its supply rises, inference leaving the rented data center for the owned desk, a single chokepoint contested at once by states, rivals, and the chip designers routing around it. Every transformation has a breaking point. A chokepoint can strangle everything that depends on it... or teach the thing it constrains every route around it that no single hand can ever close again.

AI Releases & Advancements

New today

  • JetBrains: Open-sourced Mellum 2, a 12B MoE coding model with 2.5B active parameters (64 experts, 8 active per token); ships six variants (Base, Instruct, Thinking, and SFT editions) under Apache 2.0 on Hugging Face, with a Thinking variant that produces explicit reasoning traces for multi-step agentic tasks. (JetBrains AI Blog)

Other recent releases

  • MiniMax: Released MiniMax M3, a natively multimodal model with 1M-token context powered by a new MiniMax Sparse Attention (MSA) architecture; supports image and video input and computer-use; available via API in M3 and M3-highspeed variants with open-source release planned on Hugging Face. (MiniMax Blog)
  • NVIDIA: Released Cosmos 3, an open physical AI foundation model built on a mixture-of-transformers architecture that combines vision reasoning, world generation, and action generation in a single system; ships as Cosmos 3 Super and Cosmos 3 Nano on Hugging Face with Diffusers integration, post-training scripts, and open synthetic data generation datasets. (NVIDIA Newsroom)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more