Anthropic got screwed

Anthropic faces scrutiny over its deal with Google, as critics argue the AI startup gave up too much control and future upside in exchange for cloud credits and funding. The arrangement, which includes a significant revenue share and Google's right to use Anthropic's technology, has raised questions about whether the company compromised its independence.

Anthropic got screwed Today, we're talking about Bending Spoons and Lime's IPOs; Etched's debut, Meta turning into a neocloud, and whether Anthropic got screwed. Welcome to . Cautious Optimism https://www.cautiousoptimism.news/ , a newsletter on tech, business, and power. Modestly upbeat Wednesday. Today, we’re talking about Bending Spoons and Lime ‘s IPOs; Etched’s debut, Meta turning into a neocloud, and whether Anthropic got screwed. To work — Alex 📈 Trending Up Privacy-first AI services https://techcrunch.com/2026/07/01/venice-ai-becomes-a-unicorn-with-65m-series-a-as-its-privacy-first-ai-platform-takes-off/ … corruption and office abuse https://www.cnbc.com/2026/06/30/trump-financial-disclosure-released.html … letting staff cash out early https://techcrunch.com/2026/06/30/wayve-launches-85m-employee-tender-offer-at-8-5b-valuation/ … smuggling chips to China https://www.bloomberg.com/news/articles/2026-07-01/taiwan-detains-super-micro-workers-probing-smuggling-of-nvidia-chips-to-china … British spine https://mediagazer.com/ a260630p8 … Claude Sonnet 5 https://www.anthropic.com/news/claude-sonnet-5 … hearts and minds https://starlink.com/support/article/7030ecc8-fae9-8765-bc27-51d0bd88e51a … AI discounts https://thenewstack.io/claude-sonnet-5-launch/ … fusion power https://techcrunch.com/2026/06/30/realta-fusion-generates-electricity-directly-from-a-fusion-reaction-an-apparent-first/ ? … China-Russia relations https://www.reuters.com/business/aerospace-defense/russia-approved-secret-china-military-training-top-level-sources-say-2026-07-01/ Bending Spoons’ IPO: Italian mini-conglomerate Bending Spoons priced its IPO at $29 per share https://www.renaissancecapital.com/IPO-Center/News/120177/Italy-based-tech-acquirer-Bending-Spoons-prices-US-IPO-at-$29-above-the-ran , above its proposed range of $26 to $28 per share https://www.sec.gov/Archives/edgar/data/2004711/000110465926076487/tm2613674-10 f1.htm . The company, which owns Vimeo , AOL , Evernote and several other tech companies, is valued at $18.5 billion, or $19.5 billion if you’re into more diluted share counts at its IPO price. - Bending Spoons was last valued at $11 billion in 2025 https://www.businesswire.com/news/home/20251030216227/en/Bending-Spoons-Raises-%24710M-for-Continued-Investment-and-Growth , making its IPO a massive win for all its investors, including those who put in the final checks. Bending Spoons is an odd duck. Instead of chasing AI revenues, it buys tech businesses past their prime, lowers their cost base, and prints cash to keep its acquisition machine running. It’s proved to be a solid plan: The company has a lot going for it financially https://www.cautiousoptimism.news/bending-spoons-is-the-anti-ai-play/ , it’s growing fast, and in Q1 2026, reported a strong operating profit and real net income. - How much interest public investors will have in Bending Spoons is up for debate. More tomorrow when we have more trading data. Lime’s IPO: American e-scooter company Lime priced its IPO at $25 per share https://www.renaissancecapital.com/IPO-Center/News/120184/E-scooter-rental-platform-Lime-prices-IPO-at-the-$25-midpoint , the midpoint of its proposed $24 to $26 per-share range. Lime is valued at about $1.6 billion in its IPO, or $1.8 billion if, again, you prefer a more diluted total share count. - Lime once had a valuation of $2.4 billion https://www.businessinsider.com/lime-raising-another-300-million-valuation-hits-24-billion-2019-2 , though the pandemic wrecked its business and early e-scooter economics failed to bear out as anticipated. - Still, after seeing its valuation collapse to around $510 million in 2020 https://www.pymnts.com/news/investment-tracker/2020/uber-reportedly-considering-stake-in-lime/ , the IPO has proved a win. Well done. Etched Startups are bets. They take a perspective on the market and go whole-hog after it. Sometimes, they are correct Uber , Figma , etc. and sometimes they are not. But even with that framework, it’s somewhat rare to see a startup make only one bet. Starcloud is a good example. The company can build lots of data centers in space as long as launch costs keep falling. If SpaceX’s Starship rocket works out, Starcloud may succeed. If Starship fails, and launch costs do not come down materially in the coming years, Starcloud won’t. Etched, another of my favorite startups, just said that it has raised $800 million to date https://x.com/Etched/status/2071972062202343590 , and has made massive progress on its chips. The company is not building flexible silicon that can be pointed at a variety of tasks; instead, it’s making transformer-specific ASICS — chips that only handle inferencing tasks. If the industry had moved on from transformer-based AI architectures before Etched could get to market, it would have been dead. Thankfully for the startup and its backers, the world still runs on transformers, which means it’s coming to market exactly when inference costs are spiraling and making AI customers annoyed. - Etched’s chips are not the only cost-cutting mechanism the inference market will soon enjoy: OpenAI has reportedly found some powerful mechanisms https://www.theinformation.com/newsletters/ai-agenda/openai-discovers-new-way-cut-inference-costs-half for reducing its own inference bills, and is racing to bring its own chips to market. Etched claims https://www.etched.com/ it has $1 billion in orders, and that it’s had a breakthrough in memory management. Let’s see who their customer base is and how quickly Etched can bring better economics to neoclouds and hyperscalers alike. A fun question: Is there a VC firm that has bet both on Etched and non-transformer-based AI models? 📉 https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html Trending Down 📉 https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html US-EU tech relations https://www.politico.eu/article/brussels-claps-back-trumps-tech-threats/ … Payroll growth https://www.cnbc.com/2026/07/01/private-payrolls-rose-by-98000-in-june-less-than-expected-adp-reports.html ? … Ukrainian war coffers https://www.politico.eu/article/ukraine-war-initiative-defense-kyiv-funding-defense-minister/ … CNN https://www.status.news/p/paula-reid-cnn-paramount-bari-weiss-david-ellison ? … India-Pakistan water relations https://apnews.com/article/pakistan-wars-india-water-weaponization-indus-1f8c5fff1fccdfec03192811e80ea6b7 … Chinese manufacturing growth https://www.wsj.com/economy/china-manufacturing-gauge-shows-slower-growth-in-activity-e58f6694 ? Working for tech companies: Microsoft is reportedly considering laying off as much as 2.5% of its staff, per Business Insider https://www.businessinsider.com/microsoft-job-cuts-layoffs-sales-consulting-2026-6 . The company has about 220,000 employees, so the cut should affect more than 5,000 jobs if it comes. - Layoffs in the tech industry aren’t as bad as they were during the pandemic, and are smaller in scale than during the great “ ” era of late 2022 and early 2023. But seeing the wealthiest tech companies kick employees out while thriving in the AI era doesn’t spark much joy. Oh Shit, Profitability https://layoffs.fyi/ - I wonder if layoffs are self-defeating: By shedding staff, aren’t major tech companies simply seeding new cohorts of startups with talent and, therefore, lowering the odds of their own long-term success? Compute prices: Meta is spending tectonic amounts to scale its compute and build competitive AI models. But for whom? The company keeps talking about how its AI work helps it generate more advertising revenue, but is that enough to pencil out hundreds of billions of dollars in investments? Meta doesn’t think the answer is a strong “yes.” Per Bloomberg https://www.bloomberg.com/news/articles/2026-07-01/meta-is-building-a-cloud-business-to-sell-excess-ai-compute , the company plans to sell access to both its new and upcoming AI models, and also offer raw compute capacity. In short, Meta is going to ape Amazon and build its own AWS after stacking up piles of compute for its own use. On one hand, the global technology industry is short of compute, so people will be glad to rent Meta’s GPU clusters. On the other hand, it’s not clear how much demand Meta’s new AI models will pull. - On the theoretical third hand, SpaceX has turned its compute capacity into a money-printing machine by renting it to Anthropic, Alphabet, Reflection AI, and others. Perhaps Meta is being wily in its copying after all. Anthropic drew the short straw Last night, Anthropic said the U.S. government had lifted export controls on the Mythos 5 AI model and its Fable 5 variant. After weeks of purgatory, Anthropic’s shiniest toys are back. With restrictions, however. - Claude Platform, Claude.ai, Claude Code, and Claude Cowork users can access Fable 5 starting today. - Claude Pro, Max, Team, and certain enterprise customers will receive access to Fable 5 for “up to 50% of weekly usage limits through July 7, after which it will be available via usage credits.” - Access via hyperscaler partners AWS, etc. will come as quickly as the company can manage. - Access to Mythos 5 will be restored to some US groups on June 26. - Anthropic is still working to coordinate “with the government to expand access to the broader set of domestic and international partners in the Glasswing program” for Mythos 5. Here’s the rundown of what allowed Anthropic to return its models to market: - After reports of certain jailbreaking activity with Fable 5 more on that below , Anthropic worked with the US government to train an “improved safety classifier that targets and blocks the behavior described in the report.” The company will tell users when a query to Fable 5 is rejected, and route it back to Opus 4.8, a less performant and cheaper model. - Anthropic writes that its new safety classifier “comes at the cost of flagging benign requests more often during routine coding and debugging tasks.” Amazon was widely named https://www.cautiousoptimism.news/the-anthropic-fable-mess-explained/ as the voice that caused the Trump administration to issue the unprecedented export controls on the two models. Anthropic confirmed as much in its post emphasis added : The export control directive on June 12 came after the government became aware of a report in which Amazon researchers had found a method of bypassing Fable 5’s safeguards: prompting it so that it identified a number of software vulnerabilities. In one case, the model produced code demonstrating how the relevant vulnerability could be exploited.Over the past two weeks, we have worked closely with the government and other partners, including Amazon, to review the report and evidence. Our testing confirmed that many less capable models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could identify the same vulnerabilities as Fable 5 did in the report.When it came to the demonstration of how to exploit the single vulnerability, every model we tested could produce the same demonstration as Fable 5 including Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7 . Importantly, the reported technique did not expose any unique Mythos-level cyber capabilities. The behavior reflected a borderline case for Fable 5’s safeguards—as we will explain below, there are some tasks that are unlikely to be dangerous but are nonetheless blocked by the safeguards out of an abundance of caution. The reported technique allowed access to one such behavior, but it only involved routine defensive cybersecurity work. My take on the entire saga is that the most security-minded AI lab made a model so strong that the government decided to forgo its traditional laissez faire approach to technology regulation, and quickly took control of the industry. Then, after overindexing on an interesting but not obviously dangerous report, the administration took the nation’s most powerful AI models offline exactly when open-weight models from China were demonstrating performance gains. Anthropic then had to refine the security work it was already doing to satisfy the very people who had argued against its approach to AI safety, and prove that its AI models were safe enough. Meanwhile, the technology conversation has morphed from “ How can I get my hands on the smartest AI?” to “ Hey, this Chinese GLM-5.2 model is pretty good “ Anthropic got screwed, in other words. If the government had listened to the industry earlier, we might have had precise frameworks, guidelines and rules in place for highly performant AI models and the risks they may pose. Instead, we had diddly when the government decided to act. We’re now stuck sifting through after-action reports, trying to parse what’s allowed and what’s not. What’s particularly gross are boot-licking tweets from members of the venture and startup industry praising Lutnick https://x.com/howardlutnick/status/2072100729603452965 and Wiles https://x.com/SusieWiles47/status/2072099604481335711 for allowing tech companies to do what they were already doing. - The tech-right’s praise of the government intervening in the industry’s matters boggles the mind, but it’s worth remembering how vulnerable to flattery the wealthy and influential are. The government’s hot-and-cold approach to AI regulation is also slowing down OpenAI, whose GPT-5.6 model family is still on its way to market. Meanwhile, a Chinese food delivery company has released an AI model that is exploding in popularity across the world https://www.longcatai.org/models/longcat-2 . This is the time to keep accelerating with an eye towards clear, reasonable safety rules, not the time to decelerate while making up onerous, ad hoc regulations. Instead of giving shares in their businesses to the government, Anthropic and OpenAI investors should send them a bill for lost revenue and mind-share.