Today we cover Google's new model, ponder why Databricks has leverage over its investors, and ask if OpenAI and Anthropic are big enough for their britches. #
- Welcome to *. Cautious Optimism, a newsletter on tech, business, and power. Modestly upbeat
Friday. I’m heading out on a poker trip shortly, so today’s newsletter will be slightly tömör. संक्षिप्त, if you will. To work! — Alex
📈Trending Up:Valuation haircuts…data center PR…a return to the steam-era…my lack of shock…Nubank, after earnings…Apple, the Chinese AI lab…consumer spending📉Trending Down:Legal gambling…legal gambling…Japan-Russia relations…JD.com, after earnings…personal privacy…Chinese drones in America
The Rundown #
Google makes a return: Around the same time our take on how Google had dropped the ball went live yesterday, the company released a new AI model. Talk about timing.
The new Gemini 3.7 Flash appears competitive with Claude Sonnet 5 at a price (post-discount) of $1.50/$7.50 per million input/output tokens.
Despite the oft-discussed delays in releasing its next ‘Pro’ model, demand for Google’s models has remained quite stable, per some third-party metrics. And OpenRouter is offering 3.7 Flash for 50% of its introductory price, similar to the discounts the model router offered for OpenAI’s GPT-5.6 Luna and Terra models. Goddamn, I love competition.
Model speed is becoming a critical vector: For my uses, current SOTA AI models at high reasoning levels provide all the capability I can use, at a cost (thank you, subscription pricing) that feels more than fair.
But I’ve found that I could use more speed, and others online seem to agree. AI labs have listened. Not only are Gemini 3.7 Flash and Nvidia’s recent Nemotron 3.5 Lightning the fastest small models on the market, per AA, but OpenAI has announced a new ‘Ultrafast‘ mode for GPT-5.6 Sol that makes the model up to 14x faster than its standard setting.
- Notably, Cerebras hintedat the new Ultrafast mode in its earnings call, telling investors that its chip technology “serves 5.6 Sol at a speed that is 10X faster” than other providers, and that “fast tokens are in demand and command a premium at market.”
The Ultrafast mode will deliver Sol at up to 750 tokens per second, about twice the speed of Gemini 3.7 Flash, albeit at an unlisted price. Expect to pay through the nose, and for tokens/second to become as critical a metric as token efficiency and raw token costs.
Chinese weights are coming with a price tag: At launch, Kimi K3 wasn’t accompanied by its weights, and when Moonshot did release the data, it added a clause that if you wanted to serve the model commercially, you had to remit a portion of the revenue back to the company. Alibaba (Qwen 3.8 Max), and MiniMax (M3) have taken similar steps.
Today, Z.ai announced GLM-5.3, which improves on GLM-5.2 with more post-training. The new model is better, naturally, but comes with a caveat: “The model weights of GLM-5.3 will be publicly available soon in two weeks.”
I wouldn’t be shocked to see Z.ai also dropping the model weights along with a requirement to share revenue on sales of third-party inference. As discussed, the Chinese AI’s free lunch era is coming to a close.
- Why? Because Chinese AI labs want to raise huge private rounds and/or go public. That means they need to get their customers’ noses closer to the market grindstone. So, higher rates and required commercial prices are the theme of the quarter.
Databricks: As expected, Databricks has raised yet more funding. This round ($5 billion at a $190 billion valuation) was pipped to allow the data and AI giant to buy more GPUs, but now that Databricks is quasi-public, it was kind enough to provide us with the startup equivalent of an earnings report:
- Databricks’ revenue rose more than 80% in the last year, and total run-rate revenue crossed $7 billion.
- The company maintained its adjusted free cash flow generation on a trailing-twelve-month basis.
- It also saw surging demand for its Lakehouse ($1.5+ billion run rate, >100% growth rate) and Lakebase (>$100 million run rate) products.
While the foundation model labs were busy hogging the headlines, Databricks has taken advantage of the rising demand for data usage in the AI era. The company may never go public, but we can still cheer it on from the sidelines.
Are OpenAI and Anthropic big enough for their britches? #
OpenAI and Anthropic sometimes bend to sharing a nibble or two of their financial performance whenever they raise money, and promptly shut up afterwards. But with their IPOs looming, some information is leaking out:
The FT, discussing Anthropic:“Investors expect the Claude maker’s annualised revenue to be between $100bn and $120bn by the end of 2026 — using the start-up’s preferred measure, which infers full-year sales from recent performance — up by more than 10 times over the course of 2026.”Bloomberg, discussing OpenAI:“OpenAI is on track to generate annualized revenue of more than $40 billion based on its current performance, according to people familiar with the matter, roughly doubling its run rate from the end of 2025 and bolstering the company’s plans for a Wall Street debut.”
Let’s compare that to Databricks, which with its new $190 billion price tag, is valued at 27x its current run-rate revenues.
OpenAI, at the newly reported revenue run-rate of $40 billion and its last disclosed valuation of about $852 billion, would be valued at 21.3x its run rate. Meanwhile, Anthropic would be valued at 9.7x to 8.0x its run-rate revenues of $100 billion to $120 billion, given its current $965 billion price tag.
Those multiples are precisely why Anthropic’s backers think it could fetch a valuation north of $1 trillion, even as much as $2 trillion, when it eventually goes public. OpenAI appears well positioned to defend its current valuation, though it’s critical to keep in mind that trailing GAAP revenues will be much smaller than the run-rate numbers we’re discussing. In the case of Anthropic, the market is curious how the company accounts for partner revenue costs.
Regardless, OpenAI and Anthropic are together probably clocking run-rate over $100 billion today. Against an aggregate valuation (present-day) of ~$1.82 trillion, they don’t seem absurdly expensive.
- So why does Databricks have a higher revenue multiple than the leading AI labs? First, it’s entirely de-risked: Accelerating growth paired with years of positive free cash flow make for a hugely valuable company.
- Also, given that the company has had lower capital needs thanks to its business model andcash generation, it has more leverage over investors than the cash-hungry labs do today.
That reading feels accurate given their recent momentum. OpenAI’s Codex is growing like a weed, and the company’s overall run-rate revenue grew 20% in July, while enterprise revenue grew 32% in a month.
Bring on the IPOs. So long as their cash burn goes down, the AI labs have done the hard work to grow into their valuations. Hell yeah.