← Lantern AI Daily Briefing

Anthropic prices above range at $83 for today's Nasdaq debut, Gemini 4 Neon sells cost per task, OpenAI names TensorWave for 457 MW of AMD compute, and Palo Alto pays $4.5B for Applied Compute (November 19, 2026)

November 19, 2026 · 10m 43s · Listen

Seventy to seventy-eight dollars. That was the range Anthropic took on the road nine days ago. The final number came in higher, and the stock trades today. You're listening to Lantern AI Daily Briefing, and it's Thursday, November 19th. Today: Anthropic's IPO price, Google's new middle-tier model and the cost-per-task pitch behind it, the first neocloud named on OpenAI's AMD deal, and a four and a half billion dollar bet on open-weight models from a security company. Hit follow and you won't have to come looking for the next episode.

Angelina Kleimenova at Oninvest, summarizing Bloomberg's reporting:

Anthropic and its seven co-founders raised a record $107.3 billion in total through an IPO and a private placement. The company priced its shares at $83 apiece—above the initial range of $70–78, according to Bloomberg.

We promised you this number yesterday. Anthropic opened its roadshow on Tuesday, November 10th, at seventy to seventy-eight dollars. It priced at eighty-three, and per Bloomberg and CNBC, it also grew the deal from eight hundred twenty-five million shares to nine hundred eighty million. That works out to about eighty-one point three billion dollars from the offering itself, which Oninvest says beats SpaceX's seventy-five billion dollar record as the largest IPO ever. The other twenty-six billion is the private placement with Google, AMD, Microsoft and Nvidia. And the book was nearly seven times covered when it closed on Tuesday, with several single orders above ten billion dollars, according to the Lumo daily brief. So they had room to push. Upsizing by a hundred fifty-five million shares is not a company worried about demand. The valuation depends on how you count. Fully diluted, eighty-three dollars implies about one point five trillion. Counting only shares outstanding after the offering, it's about one point zero eight trillion. ANT starts trading on Nasdaq today, the offering is expected to close Friday, and underwriters hold a thirty-day option on another hundred forty-seven million shares. Now the part I'd want a raised eyebrow on. The prospectus shows thirty-three point two four billion in revenue for the first nine months, up roughly twelve hundred percent, driven mainly by Claude Code. Against that, a net loss of a hundred four billion, which Anthropic attributes largely to accounting adjustments. Largely. Every quarter from here, the public gets to check how large. And the competitive read. Oninvest notes Anthropic becomes the first frontier model developer to go public, ahead of OpenAI. The AI Daily roundup quotes Sam Altman at the DealBook summit saying, quote, we plan to be public, with no timeline. Two labs with similar structures and similar compute bills, and now one of them reports to shareholders first.

Nat Rubio-Licht, senior reporter at The Deep View, after a week at Microsoft Ignite:

If there's one overarching thread that I've seen at Microsoft Ignite over the past few days, it's that cost efficiency is quickly edging out capability as the determining factor in whether or not an enterprise decides to use a model.

That's the frame for Gemini 4 Neon, which Google released on Wednesday. It sits between Carbon and Hydrogen, the third Gemini 4 model this month. On Artificial Analysis's Intelligence Index it scores sixty-one at its xhigh setting, third overall, behind Claude Fable 5.2 at sixty-five and GPT-6.2 Astra at sixty-two. Google says it ranks first on the Artificial Analysis Coding Agent Index, at fifty-eight point six. That coding number is the one I care about, given Astra didn't move on that index this week. Price: three dollars in, fifteen out per million tokens, fifteen cents cached, one million tokens of context. Google's pitch is per task, not per token. At the default high setting, it claims thirty-two percent lower cost per task than Carbon. Rubio-Licht's critique is the price position. Neon lands right between GPT-6.2 Sol at two and ten and Claude Opus 5.5 at four and twenty. Her worry is that a middling price for comparable performance makes it hard to stand out, especially as routing services pick models automatically. There's a date in here that matters more than any chart. Saazora reports the launch promo on Argon and Carbon ends November 30th, and from Tuesday, December 1st, their rates double to four and twenty. So if you built on Argon, Neon becomes the cheaper path by default. That's not really a capability story, it's a repricing. And keep the sourcing straight. The leaderboard placements come from Artificial Analysis and Vals. The lowest-fabrication claim and the safety numbers are Google's own internal testing.

Emily Forlini, writing in Fortune's Eye on AI newsletter:

OpenAI revealed the first neocloud company associated with the AMD deal, called TensorWave. TensorWave will provide OpenAI with 457 megawatts of capacity across multiple U.S. data center sites. That’s just 7.6% of the total capacity involved in the AMD deal.

This was announced Tuesday. It's a six-year agreement for dedicated, single-tenant capacity on AMD's Helios racks, each built around seventy-two MI455X GPUs. First capacity is slated for the first half of 2027, the full deployment by early 2028. The larger AMD deal, signed in October, covers six gigawatts. Forlini's context is why this matters. A source told her OpenAI's models run almost entirely on Nvidia GPUs, while Anthropic's S-1 shows four main chip providers. OpenAI is trying to close that gap, with a week of Nvidia leadership chaos as the backdrop. TensorWave's spokesperson told her the timing was about recruiting at the Supercomputing conference. The skeptical read comes from Tyler O'Neal at SuperComputing News. Neither company disclosed the contract value, the sites, who owns the accelerators, or how it's financed. TensorWave's last raise was a three hundred fifty million dollar Series B in June at roughly one point five five billion. Its separate two hundred twenty-two megawatt lease at Fermi America's Project Matador in Texas was terminated this month, and AMD's credit support ended with it. So a company valued at one and a half billion is on the hook to deliver hundreds of megawatts, and it told Fortune it needs to roughly triple its staff. AMD also claims thirty percent more tokens per dollar than Nvidia's Vera Rubin NVL72. O'Neal flags that as unverified, and I'd go further. Show me ROCm holding up across many racks under a real training run before I believe a tokens-per-dollar slide. One more thread between these two companies. On Wednesday, OpenAI published results saying its models improved power, performance and area on AMD's flagship GPU design by up to fourteen percent, and that AMD has put some of those changes into a design sent for manufacturing. That's OpenAI's account, but the partnership is now running in both directions.

Dealroom News, on how fast this deal came together:

The price is about 38% above that valuation, set only around two weeks earlier, making this one of the fastest multibillion-dollar exits of the current AI cycle.

Palo Alto Networks announced after Wednesday's close that it will buy Applied Compute in an all-stock deal valued at four and a half billion dollars. Earlier this month, Applied Compute raised three hundred fifty million at a three point two five billion valuation, in a round led by Elad Gil with Nvidia participating. Duncan Riley at SiliconANGLE reports it was founded in 2025 by former OpenAI researchers Yash Patil, Rhythm Garg and Linden Li, and that it says annualized revenue topped a hundred million dollars, with Microsoft and DoorDash as customers. What it sells is the boring, valuable middle: post-training and inference for open-weight models on a company's own data, with an autoscaling inference engine and a model router. Honestly, that's the layer I'd want to own too. Most teams don't need a new model, they need their model tuned and served without a platform team of twelve. And this is the argument I keep making. Open versus closed is less important than who controls the fine-tuning and the inference stack. A security vendor just paid four and a half billion to control that stack. Nikesh Arora's line is that every enterprise will run a mix of open and closed models, and every one of them needs securing. Reuters, via Harshita Mary Varghese and Shania S Thomas, lays out the risks with open weights: model tampering, data leakage, misuse. Palo Alto also beat on the quarter, with revenue up thirty-nine percent to three point four five billion. It raised full-year guidance, and the stock jumped about nine percent after hours. Patil becomes co-chief AI officer once the deal closes, expected in Palo Alto's fiscal third quarter of 2027.

A lot of today ran through safety claims we could only attribute: Google's internal evaluations for Neon, and the risk profile of open weights Palo Alto now wants to secure. If you want those taken apart properly, try our sister show, AI Safety Daily. It covers AI alignment, model evaluations, emerging risks and governance, every weekday, with what changed, what the evidence says, and why it matters. It's the show that reads the system card after we've read the price sheet.

If you're on Argon, put December 1st in your calendar and run your Neon evals before then. We'll be back tomorrow, Friday, once ANT has a full day of trading behind it. Thanks for listening to Lantern AI Daily Briefing, a Lantern Podcast.