IPO Watch: Figma raises range, Firefly sets it, Ambiq on track to trade
Welcome to Cautious Optimism, a newsletter on tech, business, and power.
Tuesday! PayPal and Spotify reported earnings this morning, leading to both companies losing ground in pre-market trading. Spotify’s earnings — see Trending Down — were particularly underwhelming. Cadence swung the other direction.
It’s been some time since we held a fascism watch here at CO (we try to avoid political blow-by-blows), but I think that recent examples of the American state undermining press freedom and intellectual freedom, crossed with greater entanglement of religion and the state are worth pointing out in a business context. Add moves by the current administration to set partisan requirements to get M&A over the line, and you have a moment in which several planks of the American system are being eroded.
In happier news, Greek startup Delian Alliance Industries just raised $14 million to build a new defense prime. Think another Anduril, or Helsing. Naturally those are big boots to fill, but rearming Europe is going to be a multi-party effort. Get some gear to Ukraine, Delian, as soon as you can. — Alex
📈 Trending Up: Gun violence … Cadence, after earnings … the politicization of regulators … and we thought the IPO drought here was bad … chances of a bipartisan housing bill … AI insurance … self-driving in Abu Dhabi … MENA-French relations …
… famine in Gaza … Calo …
Startup Round of the Day: Julius, which just raised $10 million for its “AI data analyst” in a box product. Companies that could not afford their own data team, or financial analysts, just got superpowers on the cheap.
📉 Trending Down: PayPal, after earnings … the popularity of tariffs … Spotify, after yucky earnings … national efforts to combat climate change … Russia’s free hand in Ukraine? … Denmark … UnitedHealth … cybersecurity in the bedroom …
IPO Watch
Ambiq is expected to price Tuesday and trade Wednesday. Figma is expected to price Wednesday and trade Thursday. Firefly is expected to price on August eighth, and trade on the ninth. You can read our notes on Ambiq’s finances here, Figma here, and Firefly here.
As expected, Figma raised its IPO price range to kick off the week. Now targeting $30 to $32 per share instead of $25 to $28, the American design unicorn is set to recycle billions of dollars to its backers when it debuts at a valuation that reaches $18.6 billion at the mid-point of its range. From here I expect a Figma to price a dollar or two above range perhaps getting close to the psychologically-important $20 billion mark. From there, I expect retail to get a little silly.
Firefly set first terms for its own IPO to start the week, with a range of $35 to $39 per share anticipated. Given the company’s modest trailing revenue results and incredibly well-known competition, the space launch and space vehicle company is about top earn a valuation of nearly $6 billion. That’s bonkers good, and is perhaps even more bullish than Figma raising its own IPO range. There are a bevy of spacetech startups in the market. And Firefly is telling all their backers to get more greedy on the double. As I am still waiting for space trips to match my poker budget, godspeed.
Elsewhere in money-land, don’t forget that we’re hearing from Meta, Microsoft, Robinhood, ARM, and Qualcomm tomorrow. In terms of AI sentiment, Wednesday is going to be a huge day.
Google’s notes that its token processing rates have gone vertical, implying weighty AI demand from the market. Let’s see if the social giant and Redmond can beat this (Alphabet earnings call transcript source):
The growth in usage has been incredible. At I/O in May, we announced that we processed four hundred and eighty trillion monthly tokens across our surfaces. Since then, we have doubled that number, now processing over nine hundred and eighty trillion monthly tokens—a remarkable increase.
Chinese companies are charting a very different AI path
This week Z.ai dropped new open-source AI models that are not only benchmark-impressive, but come at an incredibly low price point. Z’s new GLM-4.5 models are also incredibly cheap, with the Chinese AI company saying in a release that API calls will run users as little as “USD 0.11 per million input tokens and USD 0.28 per million output tokens.”
Z.ai was formerly known as Zhipu AI. It’s one of China’s ‘AI Tiger's’ along with Moonshot, 01.AI, MiniMax, StepFun, and Baichuan Intelligence.
Of course, you can simply self-host GLM-4.5 models if you want. Here’s the HuggingFace repo, if that’s your jam.
Much hay is being hucked over Z’s new models and their low cost thanks to prior market jitters in the wake of DeepSeek’s similar move earlier this year. At that time, the rival Chinese AI company dropped a high-quality, low-cost model and folks became worried that if top-end AI models could be made on the cheap, than quite a lot of investment into data centers and the like was going to be rendered expensively moot.
Recall that Moonshot recently dropped the Kimi-K2 model family, which is also open source and excellent.
Not so. Call it Jevon’s Paradox or simple market forces, but the race to build ever-larger data centers continues, the battle to build the newest and best AI models continues, and the market dominance of American AI models continues. This time ‘round, a new good and cheap model is worrying few.
A good question is why. Partly because the DeepSeek scare failed to actually shake the AI market’s fundamentals. But also because we’re seeing a fascinating split form between China and the United States when it comes to AI technology. While the American — read: global — market is racing ahead with closed-source models, Chinese companies are leaning into open-source AI technology.
Chinese AI companies building open-source technology is the smart business move for companies inside China’s borders in reach terms. Few folks are going to want to purchase lobotomized AI technology from Chinese companies tuned to avoid annoying local censors. So, by dropping open-source technology that others can host on their own, Chinese companies stay current in the larger AI scene without getting their doors welded shut.
Bear in mind that it’s official US policy to support open-source AI.
At the same time, growth at OpenAI and Anthropic, and probably xAI and Mistral as well, indicates that corporate customers are largely willing to pay American AI vendors for API calls instead of trying to use open-source (Chinese) AI models that they self-host.
It’s just weird that the open country is building closed models while the closed country is building open models, but as every economist will tell you, people respond to incentives.
What about Meta? Isn’t it releasing open-source models? Not really, and the company may move away from its even quasi-open-source AI roots.
What about xAI? Isn’t it releasing open-source models? Not since Grok 1.
What about OpenAI? Isn’t it set to release an open model? Yes, one, and it’s late. Big fucking whoop.
Over in China, companies are grouping under a chips banner (the “Model-Chip Ecosystem Innovation Alliance”) and under the aegis of promoting AI adoption (the “Shanghai General Chamber of Commerce AI Committee”) to help bolster local chip and code prowess. It’s probably a smart move, given that even with H20 supply set to recommence in the nation, using older American chips to chase American AI companies won’t sit well with the Standing Committee.
Reuters’ coverage of a recent Chinese AI confab that included notes on the twin alliances we just discussed included the following tidbit:
Other events included Tencent's unveiling of its open-source Hunyuan3D World Model 1.0, which the company said enables users to generate interactive 3D environments through text or image prompts.
Yep.
What about other open-source efforts?
It’s a fair point. Mistral has some open-source tech out, but what I am seeing that’s more interesting are non-model AI projects built using open-source principals. Things like this, or this, or this, or this.
Simply because China’s model companies appear to be winning the open-source AI race doesn’t mean that here in the States there won’t be open-source AI technology, when considered more generally.