Anthropic's Leaked IPO Numbers, OpenAI Dots and the Avatars That Fool Humans

Big Technology host Alex Kantrowitz and Margins writer Ranjan Roy put the week's three biggest AI stories on the same table in their first Friday show of October: Anthropic's leaked IPO figures, OpenAI's always-on assistant Dots, and avatars that fool humans on video calls. The three topics look distant from each other, yet a common denominator sharpens as the show progresses; as scale grows, questions of trust and cost move to the center of every file. The November IPO schedule is likewise confirmed in the investor-calendar analysis published by LiveMint, and the market countdown looks officially underway.

The figures Reuters relayed from the confidential S-1 draft offer the clearest picture yet, because for the first time the conversation is about booked revenue and losses rather than annualized estimates. An October 14 investor day and a pre-Thanksgiving November listing suggest the process is now tied to a point-of-no-return calendar. The Morningstar assessment based on PitchBook data argues, meanwhile, that the dizzying $2 trillion valuation expectation cannot be explained by the current pace of growth, and urges investors to stay cautious.

The leaked S-1: record growth, heavy losses

Revenue that stood at just $400 million in 2024 rocketed to $4.6 billion in 2025, with annualized revenue reaching $10 billion by year end. According to the Financial Times, single-quarter sales hit $11.5 billion in the second quarter of 2026, meaning the growth curve kept steepening even after the leak. The scale of this revenue explosion is confirmed in detail in the leak analysis published by SiliconANGLE, and the two sources' numbers match each other exactly.

The loss side looks frightening at first glance, but the picture splits in two once the accounting curtain lifts: $34 billion of the $42 billion net loss is the bookkeeping reflection of valuation gains tied to raised funding. The remaining roughly $8 billion operating loss is the true cost of the business, and more than 91% of it went to compute infrastructure. In other words the company burned about $2 for every $1 it earned, yet nearly all of that fire was paid for future capacity, not rent.

The truly breathtaking line sits not under the balance sheet but above the future: a $518 billion compute infrastructure spending plan for the coming decade, with 80% of it locked in contracts that cannot be cancelled or must be paid even if unused. Commitments of $111 billion to Alphabet, $110 billion to Amazon and $31.4 billion to Microsoft sit inside this basket, alongside a $161 billion equipment lease with Broadcom. An 80-page risk section in the prospectus shows the company knows this asymmetry and warns investors openly.

The fragility on the revenue side hides in the customer table: nearly a quarter of takings come from just two customers, and the largest customers are not tied to long-term contracts. So customers can cut spending almost at will while Anthropic sits 80% locked on the spending side; flexibility runs entirely one way. Rumors that enterprise customers are turning to usage controls to rein in token bills make this concentration risk even more acute.

Why the market is on edge

This is where the show pulls the camera back from the company to the whole index, and the view is unsettling: the S&P 500 reads 13% up year to date while the median stock sits 16% below its 52-week high. By Goldman Sachs' math, half of the index's earnings-per-share growth comes from AI investment, and stripping out AI companies leaves the index 5% down since late August. When growth rests on so narrow a group's shoulders, that group stumbling shakes the entire index.

History offers the SpaceX analogy for comfort: $18.7 billion in revenue and a $4 billion loss against an IPO above $1 trillion, with a $2 trillion valuation floated at one point. But the show treats this parallel cautiously, because October's market mood is far more skeptical and Anthropic's obligation structure differs from SpaceX. The more critical point is the cascade effect : if Anthropic stumbles on payments, cloud and chip projections at Alphabet, Amazon, Microsoft, Broadcom and Nvidia wobble too, and lawyers do the talking at the rescue table.

On the competition front, Meta's enterprise push and its Muse assistant stand out as the week's surprise, with the host openly correcting his earlier error of saying Muse was open to 3 billion users. The thought experiment on the table is provocative: what do Anthropic and OpenAI do if cash-and-compute-rich Zuckerberg starts a price war? The answer is sought in enterprise reality: big companies want the best and most capable intelligence, and cheapness alone is not enough.

Always-on assistants: Dots takes the stage

Dots, unveiled at OpenAI's annual developer day, is a different species from the classic question-answer box: an always-on agent that knows the user, works in the background and advances toward goals over weeks. Per New York Times reporting it takes on jobs from restaurant bookings to money management; it works through ChatGPT, Slack, Teams, SMS and phone calls, and launched on Pro plans. The launch demo stumbled a little, but the idea is crisp: AI no longer waits to be summoned, it acts proactively.

The rival map is crowding fast, and this race's first round coincided with launch week: Meta Muse is knocking on enterprise doors, Google made its Gemini Spark move back in May, and Instinct, fresh off a $1 billion round, stands out with its calm and tidy tone. The sharpest account of the Dots-versus-Muse divide sits in the in-depth review published by Wired, which pits the two products' proactivity philosophies against each other. A measure of how fierce the rivalry is comes in the Muse comparison published by Yahoo, which lays out the price and data-policy differences plainly.

Alex Kantrowitz tested Dots with full authority on his laptop and browser, and his notes are full of concrete examples: drafting emails, finding a Drive invoice and attaching it to a Gmail draft, planning an El Salvador trip on a $100-a-night hotel budget. He liked two of three hotel picks but kept the final booking click for himself, running rental-car and hotel correspondence in his own voice through the assistant. The picture is clear: the agent gathers 90% of the work, the seal stays with the human.

The most instructive anecdote lands at the money-and-signature point: the host tried to hand over his credit card and Dots refused it, yet he managed to get Dots to sign a WebSummit speaker contract. This contradiction reveals the product's trust architecture ; brakes on monetary transactions, throttle on legal text, and the boundaries do not yet feel intuitive to the user. That sibling product Co-work refused to sign the same contract suggests these lines vary product to product even inside OpenAI.

Ranjan Roy makes a conceptual correction here and fills in the word enterprise: to him a true enterprise means a large, complex organization that passes through security reviews, not individual productivity. He argues differentiation will ultimately come down to trust , offering his own practice as evidence: he gave Instinct his credit card but not Muse, while granting Muse his corporate card login. Manus's personal agent Q under the Meta roof and TikTok's agent rumors have joined the same queue to swim in that trust pool.

Do we need frontier intelligence

The show saves its most technical debate for last, and the question is plain: do we really need the most capable model to write emails, clone voices and run multi-step errands? Ranjan points to Ramp spending reports showing the enterprise drifting from frontier to open-weight and in-house models under cost pressure; the open-model gap has shrunk from 12-18 months to 3-6. Alex counters that he has personally seen the depth difference in recent OpenAI and Anthropic models, with far sharper diagnosis on what-did-I-miss questions in writing analysis.

This debate plugs straight into the IPO story, because always-on jobs like continuous inbox scanning get expensive fast on the most capable model. The areas where frontier intelligence genuinely shows its edge are listed as deep work: DNA sequencing, 3D simulation with world models and complex graphics. If customers slide toward cheaper models while the supplier stays locked into the priciest capacity, the squeeze threatens the profitability claim in the leak.

Has the visual Turing test been passed

Tavus's live avatar Griffin is the week's most chilling demo: a lifelike face on a video call that hears, sees, talks and reacts. In the company's test, about half of participants believed they were talking to a human; lag-free naturalness and short response times made up for imperfect pixel fidelity. This near-fifty-percent deception rate is independently examined in the detailed report published by Decrypt, which lays out the test's claims in full.

The anxiety matches the excitement, and Alex does not hide it: this technology will delight scammers most, and beyond customer service its purpose remains unclear. Text-based pig-butchering and catfishing schemes moving into the video-call era, combined with the AI slop spreading across social feeds, darken the picture. The show states the rule too: first one compliant company builds it, then non-compliant clones follow, and the inevitable end is the feed overrun by Shrimp-Jesus-grade avatars.