❯ Anthropic’s draft prospectus surfaces: a $42 billion net loss last year and at least $518 billion of compute bills over the next decade

A document not yet publicReuters reviewed the draft prospectus Anthropic filed confidentially with the SEC in June and reported its contents in late September, as relayed by The Next Web and others. The filing points to a target valuation above $2 trillion, against about $965 billion in May, with a listing expected after the US midterm elections in November. Anthropic declined to comment, and the prospectus has not been made public.

$34 billion of the loss is accountingAnthropic develops the Claude models. The draft shows 2025 revenue of nearly $4.6 billion, 12 times the previous year, and a net loss of $42 billion. About $34 billion of that was an accounting charge reflecting a rise in the estimated value of financing that could later convert into shares. The operating loss was $8.06 billion, and compute and infrastructure cost $7.33 billion, three times the 2024 figure.

Operating profit by the second quarterAccording to Fortune, Anthropic’s revenue was $4.73 billion in the first quarter of this year and $11.5 billion in the second, when it was profitable on an operating basis for the second consecutive quarter. It held $20.28 billion in cash at the end of 2025. Nearly a quarter of revenue came from two unnamed customers, and many large customers have no long-term contracts. The seven co-founders keep 50.1% of the votes through a limited liability company.

Eighty percent cannot be cancelledThe real pressure sits on the liability side. InvestmentNews relays that the company expects to spend at least $518 billion on infrastructure over a decade, about 80% of it non-cancellable: $111.1 billion with Google, $110 billion with Amazon and $31.4 billion with Microsoft, plus about $161.2 billion of equipment leases tied to Broadcom. If revenue slows, these fixed costs do not shrink, so anyone buying the stock takes on this long-term contract too.

▪ SIGNALThe prospectus reduces a model company’s business to arithmetic: revenue can grow twelvefold in a year and still has to outrun a ten-year compute bill that is 80% irrevocable.

❯ Broadcom’s bank group starts syndicating $60 billion of AI chip debt, giving Anthropic’s compute leases their funders

$60 billion looks for buyersAccording to Bloomberg on October 2, citing people familiar with the matter, Broadcom’s Wall Street bank group has begun gathering about $60 billion of AI chip financing to benefit Anthropic and other companies. The banks are preparing to send investors syndication letters for a $42 billion Class A senior secured portion. The deal has not been formally announced, and pricing, maturity and closing date have not been disclosed.

Why a chipmaker arranges debtBroadcom designs custom AI chips and networking equipment for large customers. According to The Next Web in August, under this structure Anthropic does not buy the chips directly. Investors pay for them and lease the hardware to the company. Broadcom would also backstop part of the senior layer so those securities can earn investment-grade ratings and lower borrowing costs.

From $35 billion in June to nowBroadcom, Apollo and Blackstone formed a partnership in June whose first transaction was $35 billion to expand Anthropic’s compute. In August Broadcom was reported to be in talks for more than $60 billion of new debt. This week’s plan is a $42 billion senior layer plus an $18 billion junior layer led by Blackstone, which is taking $9 billion itself and syndicating the rest (FourWeekMBA, relaying Bloomberg).

Creditors bet on Anthropic's behalfThe $42 billion figure matches the ceiling of the loan Broadcom agreed to make to Anthropic in the draft prospectus, but neither report confirms they are the same instrument. What is clear is that bond investors advance the money for the chips and are repaid from Anthropic’s future lease payments. Their question is not whether Broadcom’s chips are good, but whether Anthropic’s revenue can cover the rent for years on end.

▪ SIGNALWhen the supplier guarantees, private credit pays and the customer only pays rent, selling AI chips has become a credit business, and an order is only as good as the borrower’s cash flow.

❯ Smart ring maker Oura postpones its $2.2 billion IPO and chooses to wait despite orders four times the shares on offer

Halted before pricingOura announced on September 29 that it was postponing its initial public offering, citing uncertainty in the US IPO market, and gave no new date. According to a report carried by Gulf News, it had planned to sell 50 million shares on Nasdaq at $40 to $44 each, raising up to $2.2 billion at a fully diluted valuation of about $15 billion at the top of the range, and orders came to roughly four times the shares available.

Rings, and above all subscriptionsOura was founded in Finland in 2013 and makes smart rings worn on the finger that track sleep, recovery, activity and stress. According to MedTech Dive, it runs on a subscription model with 5.7 million paying members, had revenue of $907.9 million in fiscal 2025, is profitable, and expects revenue to grow 90% in 2026. Its last private round raised more than $900 million at a valuation of about $11 billion.

Most of the shares came from insidersOf the 50 million shares in the offering, the company was issuing only 13.5 million. The other 36.5 million were being sold by existing shareholders. Holtec Nuclear and Bamboo Insurance Services also delayed listings in the same period. CEO Tom Hale said the company wants to deliver an extraordinary IPO for employees and investors and that “we have the luxury of choosing our moment.”

Public markets do not take everythingThe Information argues that Oura’s delay shows a widening divide in venture exits: large acquisitions are getting done, while public investors push back on IPO valuations and shareholder sales. If a profitable, fast-growing company has to wait, those queuing behind it will find pricing harder still.

▪ SIGNALPulling a deal that was four times oversubscribed shows the obstacle to an IPO is not whether buyers exist, but at what price and who gets to cash out first.

❯ Personal AI assistant Instinct raises $1 billion as its valuation goes from $2.5 billion to $10 billion in a month

An assistant that makes the call for youInstinct is a San Francisco company building a personal AI assistant. Users text or call it with a task and it plans the trip, orders the groceries or books the tickets. According to OfficeChai, it connects to the user’s email, calendar and messaging apps, is built mainly on open-weight models, is invite-only and free, and has announced no pricing.

A 23-year-old founder and 14 employeesFounder and CEO Noah Shinn is 23. He did machine learning research at Northeastern and MIT, dropped out in 2023, wrote the Reflexion paper on language agents and then worked as a research scientist at Sierra. The company has just 14 employees. Runtime Wire relays that The Information reported in mid-September that it had more than 100,000 users.

Four rounds in five monthsOn September 28, Instinct closed a $1 billion Series C at a $10 billion valuation from Sequoia Capital, Benchmark and Coatue. Its earlier valuations were about $50 million, then $500 million at a Series A in early August (raising $75 million), then $2.5 billion at a Series B on August 26 (raising $250 million).

$10 billion without revenueInstinct has disclosed no revenue or paying-user figures, and Runtime Wire notes that capacity strain and questions about data exposure came up during testing. It faces Meta’s Muse and the assistant OpenAI is building. Investors paying $10 billion are betting it can be first into users’ inboxes and calendars. Once those permissions are handed over, switching assistants is costly, but only if it avoids a privacy failure.

▪ SIGNALA free, invite-only company quadrupling in value within a month means capital is pricing a position, namely who first holds a user’s full account access, not revenue.

❯ Robotics software company FieldAI signs a term sheet for $700 million, lifting its valuation to $10 billion

Robots that need no mapFieldAI was founded in Irvine, California, in 2023 and makes only software, not robots. According to SiliconANGLE, its models let robots operate in places with no pre-built map, no GPS and no internet connection, and adjust their behavior to risk, for example slowing down when lighting fails. It also builds digital twins of sites from camera and lidar data.

Customers on building sites and in power plantsIts customers are in construction, energy and the public sector. In March it partnered with Boston Dynamics so that the Spot robot dog could use its software for industrial equipment inspection. According to TechCrunch last August, it had then disclosed $405 million in total funding from investors including Bezos Expeditions, Temasek, Khosla Ventures and Intel Capital.

Term sheet signed, not closedBusiness Insider reported on October 2 that FieldAI is raising $700 million at a $10 billion valuation, five times its valuation last August. What has been signed is a nonbinding term sheet. Existing backers are expected to take part and new investors were not disclosed. Revenue and signed contracts together exceed $135 million across more than 30 customers, up from $100 million in June.

Software valued above whole robotsThe Next Web compares it with fellow robotics software companies Physical Intelligence, valued at about $11 billion, and Skild AI, at about $14 billion, while Europe’s largest robotics rounds this year all went to hardware, the biggest being NEURA at about $7 billion. Robot makers therefore face a choice: pay to build the brain themselves, or buy it from these software companies and give up part of the margin.

▪ SIGNALA company that builds no robots is worth more than companies that do, because capital is betting bodies will converge and software that works across bodies is the scarce part.

❯ General Intuition raises $220 million at a $6.2 billion valuation as its models trained on gameplay footage reach customers

Learning actions from player recordingsGeneral Intuition does not build models that process text. It builds foundation models that learn actions, with the aim of letting AI act in real time in environments it has not seen. Its training data comes from the gameplay video platform Medal, where uploads are paired with the player’s inputs. According to Pulse 2.0, Medal is on track for about 3 billion uploaded videos a year.

Spun out of Medal a year agoThe company was spun out of Medal in 2025, and its CEO and co-founder is Pim de Witte. According to Runtime Wire, it raised a $133.7 million seed round in October 2025 led by Khosla Ventures and General Catalyst, then $320 million on June 25 this year at a $2.3 billion valuation. It has shown a model called MIRA that generates four-player Rocket League games.

Nearly triple in three monthsOn September 29, the company announced it had raised $220 million at a $6.2 billion valuation from Valor Equity Partners, Atreides, Seven Seven Six, Point72, Khosla Ventures and General Catalyst. It also began making its models available to customers, with partners currently gaining access through an early-access list. The money will go to more training, compute and researchers.

One step from games to realityWhat robotics companies lack most is action-labeled data, and General Intuition says leading robotics models train on less than 1% of the action data it can reach through Medal. If skills learned in games transfer, robot makers gain a data source far cheaper than collecting on real machines. Runtime Wire cautions that such transfer to physical robots at scale remains unproven.

▪ SIGNALText models grew up on the internet’s text, and companies building action models are looking for their own internet. For now the closest thing is billions of gameplay clips with the inputs recorded.

❯ Rental-housing AI company EliseAI raises $350 million at a $4 billion valuation and takes the same system into clinics

Answering messages for landlordsEliseAI was founded in 2017 by Minna Song and Tony Stoyanov and sells AI software to apartment operators that automates administrative work such as tenant inquiries, lease management and tour scheduling. According to Crypto Briefing, its platform is used by one in six US apartments.

Over $200 million in annual revenueThe company passed $100 million in annual recurring revenue in 2025 and exceeded $200 million by the middle of this year. It has extended the same approach to healthcare providers, handling appointment scheduling and patient intake. In September it launched a new product called Apollo, described as a cross-platform AI tool.

Led by a16z and BessemerOn September 29, EliseAI announced a $350 million round at a $4 billion valuation, led by Andreessen Horowitz and Bessemer Venture Partners, with Ontario Teachers’ Pension Plan, Sapphire Ventures and Navitas Capital participating. According to AI Weekly, it was valued at $2.2 billion in August 2025. The money will go to new products, larger engineering and sales teams, and a second engineering hub in San Francisco.

The math is easy in a verticalProperty managers and clinics are buying saved front-desk labor: inquiries and scheduling are high-volume, rule-bound work. EliseAI’s valuation is under 20 times its annual recurring revenue and rests on revenue it already has. What it must show next is that an approach proven in rental housing sells just as well to healthcare providers.

▪ SIGNALIn the same week, a vertical AI company with revenue was priced at under 20 times sales and a general assistant with none at $10 billion, so capital’s tolerance for the two stories differs by an order of magnitude.

❯ GPU cloud provider GMI Cloud secures $223 million in equity and a $445 million credit facility to expand compute in the US and Asia-Pacific

Renting GPUs to model companiesGMI Cloud was founded in 2021 and is based in Mountain View, California. Its founder and CEO is Alex Yeh. It builds GPU clusters, rents compute to AI teams and offers inference services, with customers including Fireworks, OpenRouter, Reflection, Cartesia and Trend Micro. According to Pulse 2.0, it processes about 4 trillion tokens a week.

Contracted revenue up ninefold in nine monthsThe company says its contracted annual recurring revenue exceeds $600 million, nine times the level at the end of 2025, while live revenue has grown 4.5 times. Its data centers are in the United States, Taiwan and elsewhere in Asia-Pacific and run Nvidia GB200 and GB300 NVL72 systems. Yeh describes its place in Taiwan’s supply chain as how it keeps delivery promises.

Equity and debt are separateOn September 30, GMI Cloud announced $668 million in financing. Of that, $223 million is Series B equity led by San Francisco investment firm ARCHIV, with Nvidia, DSC Investment, Trend Micro, KB Investment, Kyobo Life and KT Corporation participating. The other $445 million is a credit facility led by CTBC. A facility is a ceiling on what can be borrowed, and the report does not say how much has been drawn.

The gap between signed and liveContracted revenue far above live revenue means customer orders are running ahead of machine deliveries. The facility lets it buy GPUs before collecting rent, but interest starts on the day money is drawn. The lag between hardware arriving and customers actually using it decides whether this debt is an accelerator or a burden.

▪ SIGNALSecond-tier GPU clouds are being financed more and more like real estate: a little equity, a lot of secured lending, and pre-leases holding up the expansion.

❯ onsemi changes its Synaptics acquisition from a roughly $7 billion stock deal to about $5.7 billion in cash

Stock becomes cashonsemi and Synaptics announced an amended merger agreement on October 1. According to their SEC filing, onsemi will acquire Synaptics for $123 per share in cash, about $5.7 billion in total. The original agreement signed on June 25 was an all-stock transaction valued at about $7 billion.

Power chips plus edge AIonsemi makes power and sensing chips for automotive, industrial and AI data center customers. Synaptics provides edge AI solutions including embedded compute, wireless connectivity and multimodal sensing. In another filing, the companies say the combined business will serve the edge AI and “physical AI” markets.

After a third-party approachThe amendment followed an unsolicited, non-binding competing proposal to Synaptics from a third party, whose identity and terms were not disclosed. Synaptics’ board unanimously determined that the amended onsemi deal remains in shareholders’ best interests. Closing is expected in mid-2027, subject to Synaptics shareholder and regulatory approvals. The US Federal Trade Commission has cleared it, and other jurisdictions are still reviewing.

Who carries the share-price riskIn a stock deal, what Synaptics shareholders finally receive depends on onsemi’s share price. With cash, the amount is fixed at $123 a share. onsemi will pay with cash on hand plus committed debt financing from Morgan Stanley, says the deal will be immediately accretive to non-GAAP earnings per share, and expects $200 million of annual synergies.

▪ SIGNALIn the consolidation of edge AI chips, the seller prefers certain cash to the buyer’s stock, which is itself a statement about chip-stock valuations.

❯ Supabase raises another $150 million and acquires Turso to prepare for AI agents creating databases in bulk

A back-end toolkit for developersSupabase offers a development platform centered on the Postgres database, with authentication, file storage, real-time subscriptions and vector search, so developers can build apps without assembling their own back end. According to Runtime Wire, the platform has 13 million developers and adds 4 million databases a month, 70% of them created by AI agents or AI tools.

It raised $500 million in JuneIn June, Supabase closed a $500 million Series F at a $10 billion pre-money valuation, led by Singapore’s GIC. According to Trending Topics, it had about 10 million developers at the time, and its database count was growing 600% year on year.

Part of the new money cashes out staffThe $150 million announced this week is also led by GIC, with CapitalG, the growth fund of Google’s parent, IronArc and SquarePeg participating. The valuation for this round was not disclosed. Part of it lets employees sell shares they hold, and the rest goes to new features for AI agents. The company also introduced Supabase Compute, hosted sandboxes for long-running agents.

One database per agentSupabase also announced it is acquiring Turso for an undisclosed price. The deal is agreed but not closed. Turso is an open-source, SQLite-compatible database written in Rust that treats a database as a lightweight file instead of a standing server process and can be created on demand. Most of the 4 million databases Supabase adds each month are made by agents, and it needs a form of database that can be created instantly and in large numbers, which is the gap Turso fills.

▪ SIGNALThe main “user” of a database is shifting from developers to agents, so products will compete less on features and more on how cheaply and quickly an instance can be created.

❯ Edge AI chip company SiMa.ai closes a $150 million Series C at a $1.45 billion valuation

Running AI on the robot itselfSiMa.ai was founded in San Jose in 2018. It designs chips and the software to go with them so that robots, drones and cars can run AI locally instead of sending data back to the cloud. According to The Next Web, it positions itself as a cheaper, lower-power option to Nvidia’s GPUs, and its customers and partners include Bosch, Emerson, Micron, Synopsys and TRUMPF.

Revenue quadrupled in a yearFounder and CEO Krishna Rangasayee was previously chief operating officer at chipmaker Groq. The company says revenue quadrupled between 2024 and 2025 without disclosing the amount. Target uses include drones, humanoid robots, driver assistance and car cockpits.

Led by Fidelity and AmplifyThe Series C announced on September 28 totals $150 million at a $1.45 billion valuation, bringing total funding to $500 million. Fidelity and Amplify led, with existing investors including Dell Technologies Capital, Maverick Capital and Point72, and new investors AllianceBernstein, Baron Capital, J.P. Morgan and the State of Michigan. The money will scale its developer software Palette Neat and fund new hardware planned for the first half of 2028.

An alternative to NvidiaSiMa.ai competes on power draw and price, and says its software cuts deployment time from months to days. Its difficulty is that the next generation of hardware does not arrive until 2028, and customers may be locked into Nvidia’s edge products in the meantime.

▪ SIGNALCapital is buying both the software brains of robots and the chips that run them, which shows nobody is sure which layer of physical AI will keep the profit.

❯ Optical interconnect company CScale exits stealth with a $145 million Series C, with Nvidia and Intel Capital investing

Lasers fail, compute should notCScale was founded in Palo Alto in 2023 and develops optical interconnect for AI data centers, using light to link thousands of accelerators into a single system. According to The Next Web, its pitch is to contain optical failures so that one failed laser does not interrupt the whole job. CEO Martin Lund puts it this way: “Lasers will fail. Compute shouldn’t.”

Two veteransFounder and CTO Sanjai Kohli co-founded the GPS chip company SiRF. CEO Lund previously led Cisco’s Common Hardware Group. The company has about 85 employees and has been backed by Sutter Hill Ventures since its inception.

First strategic investorsOn September 30, CScale came out of stealth and announced a $145 million Series C led by Atreides Management and co-led by Valor Equity Partners and Premji Invest, with Sutter Hill and Maverick Silicon participating. Nvidia and Intel Capital joined as its first strategic investors. Total funding is $188 million. The company disclosed neither customers nor a product timeline.

A crowded fieldThe same area has seen several large rounds in recent months: Lumilens raised $700 million in August, Celero closed a $275 million Series C in September, and Eliyan closed a $145 million Series C in July. How much any of them sells depends on whose design system makers such as Nvidia adopt, and Nvidia investing in CScale is not the same as adopting its product.

▪ SIGNALChip giants investing in several interconnect startups at once are buying options, and for a startup, winning an investment and winning an order are different things.

❯ Volantis raises an $88 million Series A to link AI chips to pooled memory with light

Memory out of reachVolantis is based in San Francisco and wants to solve the problem of AI chips being unable to reach enough memory. According to Runtime Wire, it uses optical waveguides and micro lasers to connect processors to shared memory chips. Electrical signals travel only millimeters inside a package, while its optical links can travel more than 200 millimeters at under one picojoule per bit, so one processor can reach far more memory.

People from Micron and Ayar LabsCEO Tapabrata Ghosh previously co-founded Vathys. CTO Roy Meade led Micron’s HBM (high-bandwidth memory) program and was a vice president at optical interconnect company Ayar Labs. The company emerged from stealth in June 2025 with a $9 million seed round.

A-1 arrives in 2027The Series A announced on October 1 totals $88 million, co-led by Lachy Groom and Abstract Ventures, with John Doerr, VXI Capital, Triatomic and Susa Ventures participating, for $97 million in total funding. Its first product, A-1, is a photonic inference system with first customer deliveries planned for 2027, targeting models of more than 20 trillion parameters and up to 10,000 tokens per second per user.

The numbers are still targetsThose performance figures describe a system that has not shipped. The chip has been taped out but not yet evaluated in silicon. If it works, AI accelerators could reach larger and cheaper pools of shared memory, which is the core story the company tells investors. Until the product is delivered, it remains a design to be proven.

▪ SIGNALThe bottleneck in AI hardware has moved from compute to memory, and startups are not trying to build better memory but to let chips use memory that is farther away and cheaper.

❯ Nebius acquires Inferize, less than a year old, to cut model cold starts from tens of minutes to seconds

Eliminating idle GPUsInferize is an Israeli company whose software shortens cold starts. A cold start is the loading time between launching an AI model and its being ready to answer requests, during which GPUs draw power without earning revenue. According to CTech, its CEO says the technology can cut loading time from tens of minutes to seconds.

Nine months, 17 peopleInferize was founded earlier this year, has 17 employees in Tel Aviv, and had never disclosed any funding. CTech reports that it raised a $10 million seed round led by TLV Partners. Founders Guy Bortnikov and Lior Gorbonos were both part of the founding team of Israeli company Granulate, which Intel acquired in 2022 for about $650 million.

Price not disclosedNebius announced the acquisition on October 1 in a filing with the SEC and did not disclose terms. CTech estimates the price at between $100 million and $150 million. Inferize’s technology and team have joined Token Factory, Nebius’s managed inference platform.

The margin in rented compute is utilizationNebius is a Nasdaq-listed AI cloud company headquartered in Amsterdam. Inference is sold by the token, customer demand rises and falls, and slow model loading means keeping extra GPUs running in reserve. Faster loading lets the same GPUs serve more requests. The filing notes that Token Factory also integrates optimization technology from earlier acquisitions Eigen AI and Clarifai.

▪ SIGNALGPU clouds have finished competing on who can buy chips and started competing on who leaves them idle least, which is why a 17-person software team is worth over $100 million.

❯ AI agent security company Reco adds $55 million as customer AT&T becomes a strategic investor

First, count the agentsReco works on agent security: helping large companies find which AI agents are running internally, under what identities, with what permissions and with access to what data. According to Pulse 2.0, its Reco Graph maps the relationships between agents and applications, data and workflows, and supports more than 280 app integrations, including OpenAI, Anthropic, Microsoft Copilot, Salesforce, ServiceNow and Workday.

Topping up after the Series BThe company’s CEO and co-founder is Ofer Klein. According to Reco’s blog, its total funding stood at $85 million after its Series B. TechCrunch describes the new money as an extension of that round.

AT&T as customer and shareholderThe round announced on September 29 totals $55 million. AT&T Ventures took part as a strategic investor, and Forestay and Quadrille Capital are new investors, bringing total funding to $140 million. AT&T is itself a Reco customer. The money will expand sales, partner channels and customer support.

More agents, messier permissionsOnce a company connects agents to systems such as Salesforce and Workday, each agent carries a set of permissions, and security teams often do not know the total. Reco cites a Grand View Research forecast that the agent security market will grow from $1.3 billion in 2025 to $17.8 billion by 2033. That is one firm’s view, but a customer as large as AT&T paying to take a stake at least shows demand has moved from trials to procurement.

▪ SIGNALCompanies are deploying agents faster than they can count them, so the first thing security vendors sell is not protection but an inventory.

❯ Nine-person Halluminate raises a $30 million Series A selling finance-task training environments to top labs

Setting finance exams for modelsHalluminate was founded in San Francisco in 2024 by Cornell computer science alumni Jerry Wu and Wyatt Marshall and has nine people. According to AI Weekly, relaying Fortune, it offers AI labs three things: sandbox environments for training agents, benchmarks that measure performance on financial tasks, and expert evaluation services.

The top score is only 51%Its Westworld Finance Diligence Bench contains 88 tasks drawn from anonymized real private-equity transactions. Across seven models tested, the highest score was 51%. Labs pay to find out where their models fall short on this kind of work and then train them in environments aimed at those gaps.

Led by Oak HC/FTThe Series A announced on October 2 totals $30 million, led by Oak HC/FT with Y Combinator, Orange Collective, Heavybit and FT Partners participating, along with angels from Anthropic, OpenAI and Meta. Total funding is $38.5 million.

Nine people, already profitableThe company says four of the top five US closed-source AI labs are paying customers, that its annualized revenue run rate is in the mid eight figures, and that it is profitable. A run rate projects a full year from the current pace and is not realized annual revenue. Having customers concentrated in a handful of labs is both its strength and its risk: if labs decide to build such environments themselves, the orders disappear together.

▪ SIGNALFor models to learn the work of bankers and private-equity analysts, what is scarce is not compute but exam questions written by people who know the job, so the small firm selling the questions made money before the financial firms using the models.

OUTLOOK

Where the money comes from and whose books it sits onRead together, this issue splits AI’s funding chain into three parts. Upstream are model companies such as Anthropic, with fast-growing revenue and ten-year compute commitments that are 80% non-cancellable. In the middle are those advancing the money: Broadcom arranging debt and GMI Cloud buying GPUs on credit, with the risk landing on creditors. Downstream is a group of startups whose valuations are rising quickly. EliseAI and Halluminate have revenue or even profit, while Instinct has not started charging. Oura’s delay shows public markets are stricter about price and insider selling. The thing to watch next is how Anthropic prices, because it will tell everyone what public markets will pay to take on the obligations along this chain.