The Week in Short
OpenAI’s annual recurring revenue is $20 billion less than news reports had indicated, spotlighting fuzzy math issues. Open-source models from the US and Europe launched to media fanfare this week and vowed to take on China, but national origin doesn’t matter much to the founders using them. Startup funding so far this year shatters 2021 records but slows in latest quarter. Singapore startup Manus raises after its forced split from Meta. Google’s AI push in the classroom proves a nightmare. Nvidia CEO Jensen Huang’s son-in-law takes on a bigger role inside the company. USV among the VC firms raising new funds. A new lawsuit over Groq’s deal with Nvidia could throw a wrench in the reverse acquihire strategy. Tech moguls get the Hollywood treatment.
The Main Item
OpenAI Wanted Its Revenue Run-Rate to Match Anthropic’s. Cue the Chaos
AI stocks got hammered Thursday after the Financial Times reported that OpenAI was spinning its annualized revenue figure to investors in order to match Anthropic, and that its ARR was actually $20 billion below a widely cited $70 billion figure.
On the one hand, it was simply a confusion about gross versus net revenue.
On the other hand, OpenAI, Anthropic and the scoop-hungry media have been complicit in the confusion. OpenAI has long been complaining that its revenue figure looks feeble compared to Anthropic's because it’s been using a net figure while Anthropic has been using gross (ie Anthropic's number includes the sales of models by partners like Amazon and Microsoft, even though the partners take a percentage of those sales).
But when the FT got its hands on OpenAI's latest revenue figure — circulated as part of the company's current funding round — the paper blew it up into a headline saying it was $20 billion below earlier reports. The story explained this, but stocks took a hit anyway. Bloomberg later reported that OpenAI is expecting to reach $70 billion annualized revenue by the end of the year — with the assumption that this is net and not gross. Investors and customers believe that Anthropic and OpenAI are running more or less neck-and-neck.
The reality here is that sophisticated investors looking at these figures should understand the gross versus net discrepancy. But Dan Primack put it best:
A lack of transparency is also at the heart of the tranched-round issue that we’ve been writing about, as Natasha Mascarenhas at Bloomberg points out:
Deals at multiple valuations are becoming normalized, and the optics are one of the reasons. As journalists we’re pro-transparency in general, and there are real risks for investors, employees, and founders alike if the true valuation figures aren’t made clear.
China Challenge
Reflection AI, Mistral & Germany’s Aleph Alpha Aim to Invent Open-Source Business Models
Until now, the policy debate over the possible risks of open-source AI models has been wrapped up in the China question. Not anymore.
American and European companies this week took aim at red-hot Chinese rivals such as DeepSeek with the announcements of three new open models: Reflection’s Beam in the US, Mistral’s Large 4 in France, and Aleph Alpha’s Kolibri in Germany. All three startups were openly positioning their new models as the ones to pick if you’re concerned about AI sovereignty.
Reflection’s new model enjoyed a blitz of coverage across Axios, Reuters, and Fortune, with the company press release explicitly saying it “advances the frontier for the Western open ecosystem.” Beam’s core claim is that it’s even more efficient on inference than Chinese model maker Z.ai’s highly touted GLM 5.2, though that isn’t Z.ai’s latest offering.
Absent government pressure or new regulations, though, the sovereignty pitch may prove a tough one. Investors we spoke with this week said few of their portfolio companies were concerned about the national origins of their open-source models at this point, other than those working with sensitive trade or government data. They’d prefer a Western model in theory, but many will gladly use a Chinese one hosted by a US inference provider like Baseten, Fireworks, or Modal.
Thus far, proponents of open source have prevailed in arguing against regulation. But the obvious risks of cutting-edge AI in the wrong hands, combined with the business threat that cheap open-source intelligence could pose to OpenAI and Anthropic, mean the question isn’t going away.
Bill Gurley, a vocal open source backer, took to Substack this week to defend open source and pitch it as an ideal AI business strategy. Open models, he argues, are more secure than proprietary systems because the wider community stress-tests them — a concept that proved true in an earlier era of computing. Any regulation based on the potential dangers of open source will likely manufacture a monopoly for frontier labs, he says. “IBM was not allowed to outlaw Dell.”
Benchmark even backed Singapore-based AI provider Manus and cashed in when it was bought by Meta, though that deal was unwound by Chinese regulators. Manus this week raised more than $500 million in fresh capital after its founders bought it back (see below).
Menlo Ventures partner Deedy Das noted that cheaper, high-performance open models are great for the application-layer companies, and not such a big threat to the frontier model-makers.
“I find it all very strange that people pit them against each other because I think there’s clearly a world where the pie grows faster and everyone grows,” he said.
Is Open Source Cheaper?
Reflection claims it can run comparable queries on a fraction of the compute needed before, so cost is certainly part of the pitch. Yet it’s not clear whether open-source models will ultimately retain a massive pricing advantage.
Another investor we spoke with said implementing open-source models can get expensive quickly, given how much time and infrastructure spending is needed to customize them for specific tasks. Outsourcing it to an inference provider can be cheaper. But the total cost is still a function of how much you call on it, and the tokens add up.
When it comes to underlying performance, the new Western rivals are comparable to the Chinese models on some benchmarks, but behind on overall capabilities.
They haven’t yet dented the rankings when it comes to demand: Top Chinese models like GLM 5.3, Kimi K3, and DeepSeek V4.1 remain the most popular on the cloud platforms Baseten and Fireworks.
And of course nothing is slowing down in China: DeepSeek is close to raising $12 billion in funding co-led by Tencent and battery company Contemporary Amperex Technology Co. per Bloomberg.
How the business model for open-source LLMs will shake out remains an open question. Low prices are good for application-layer founders, of course, but LLM training remains very expensive.
We have already seen some companies step away from this race. Databricks CEO Ali Ghodsi told Newcomer this summer that by the time Databricks released its own open-weight foundation model in 2024, he was ready to move on from such an expensive and laborious task that could only produce a product that quickly became obsolete.
Tech giant Meta is still building open-source models, though it did pivot with its first-ever closed-source release this year, Muse Spark. The new open-source startups will need to be creative and nimble to avoid a similar change of course.
AI 100
Newcomer Partners with Bessemer for AI 100
We’re thrilled to be producing this year’s AI 100 list alongside Bessemer Venture Partners to highlight the most important AI companies in Silicon Valley.
Submissions for the list are due Oct. 15. We’ll share the full list on-stage at the Cerebral Valley AI Summit on Nov. 12.
Submit the companies you want to see represented here.
Two Big Charts
Deal Value in 2026 Smashes 2021 Record, Q3 Slows to More Normal Pace
It’s been a year of massive, concentrated bets in venture, pushing dollar totals to new heights.
Deal value for US startups has already surpassed 2021’s record-breaking total by around 44%, according to the latest quarterly PitchBook-NVCA Venture Monitor report.