# China finally built an American AI model > Source: > Published: 2026-07-17 16:55:05+00:00 # China finally built an American AI model ## And: Netflix earnings, posting cringe on main, and IPOs! **Friday.** Stocks are broadly lower, with chip companies losing more ground today amid broader tech weakness. The chip selloff is strong enough that Apple retook the [market cap crown](https://companiesmarketcap.com/) from Nvidia this morning. (Pour one out for leveraged ETF holders!) Today, we’re diving into the longevity of cringe posts, IPO results (data centers, nuclear fuel), and a rundown of Moonshot’s new, impressive, and not-cheap AI model, Kimi K3. To work! **— Alex** *A few months back,*<3**CO**[reached 10,000 subscribers](https://www.cautiousoptimism.news/10000-subscribers-long-form-audio-and-opus-4-7-is-a-letdown/). That was good to see. We recently reached 10,000 daily opens, which is even better. Appreciate you all. ## 📈 Trending Up [A distinct lack of surprise](https://apnews.com/article/ice-david-brouillette-johan-guerrero-maine-shooting-dbc30d6d59e2a95fb470afc188e125c6?utm_campaign=trueAnthem%3A+New+Content+%28Feed%29&utm_medium=trueAnthem&utm_source=twitter)…[profiting from office](https://x.com/BrendanPedersen/status/2077844213559963969)…[public corruption](https://x.com/kylegriffin1/status/2077794178583437350)…[housing starts](https://x.com/FirstSquawk/status/2078095605797462172)! …[smoke-driven self-driving recalls](https://www.marketscreener.com/news/zoox-recalls-self-driving-cars-because-they-may-not-detect-smoke-ce7f51dad88bf026)…[investing in compute](https://techcrunch.com/2026/07/17/why-the-first-gpu-financiers-are-turning-to-inference-chips-in-a-400-million-deal/)…[calls for global AI](https://www.reuters.com/world/asia-pacific/chinas-xi-promotes-chinas-commitment-ai-access-speech-shanghai-conference-2026-07-17/) **Posting cringe on main:** Did Airbnb CEO Brian Chesky post a thread on Twitter discussing “real-world asset tokenization?” The cringey dreck ([preserved here](https://finance.yahoo.com/markets/crypto/articles/strangers-wont-share-home-airbnb-090833310.html)) that briefly appeared on his X profile was quickly scrubbed out, and Chesky [claimed](https://x.com/bchesky/status/2077890121232515437?ref_src=twsrc%5Etfw) his account had been compromised: “To the person who hacked my account earlier this week: thanks for all the new crypto followers.” He added that his new web3 fans would find him a “very disappointing follow.” As expected, not [everyone](https://x.com/nic_carter/status/2077898651733467419) found [the](https://x.com/bogwitch/status/2078070940210995712) hack [story](https://x.com/markvalorian/status/2077907644732334305) credible. The saga reminds me of Bessemer’s flop thread from late June. That series of posts began by discussing how GPUs were no longer the most pressing bottleneck in data center construction, as other constraints, such as power, were growing more pertinent. It felt like a missive published entire quarters too late. That thread was deleted, too, but [the mockery persists](https://x.com/sahildotcom/status/2068659558520680914). **Split screens: **As [labor’s share of GDP reaches an all-time low](https://fred.stlouisfed.org/series/LABSHPUSA156NRUG) (at least since the government started keeping records in 1950), the populace is downright depressed about the state of the economy. A [recent CNBC poll](https://www.cnbc.com/2026/07/17/economic-outlook-is-worsening-and-trump-is-getting-blamed-cnbc-survey-finds.html) found that “47% of the public are cutting back on essential items, like food and medical care, up six points from the April survey.” The same survey found that only 25% of Americans are “optimistic about the economy.” Meanwhile, the wealthiest Americans are reaching new monetary heights, and [are spending more and more of that money on elections](https://www.wsj.com/politics/elections/ken-griffin-has-poured-40-million-into-helping-republicans-win-the-midterms-f0484274). And with the stock market trading around all-time highs, it’s a great time to own assets, but the folks doing the work are less enthused. Keep the above in mind for the upcoming midterm elections, and remember that the public’s backlash against AI is partly predicated on the concern that it will automate jobs that currently feed many families, while the wealthy benefit even more. - I’m not saying that that will happen, or even that I agree with the concern. But it’s critical to keep tabs on what the average person thinks and feels, not just on what people drunk on riches post on Twitter. [📉](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) [Data center inertia](https://sourcenm.com/2026/07/16/new-mexico-land-commissioner-blocks-project-jupiter-related-pipeline-from-building-on-state-land/)? …[USA-CA relations](https://www.semafor.com/article/07/17/2026/republican-senator-goes-after-canada-for-wildfire-smoke)…[SpaceX stock](https://finance.yahoo.com/quote/SPCX/), after a[scrubbed Starship launch](https://finance.yahoo.com/markets/stocks/article/spacex-stock-drops-after-first-test-flight-since-ipo-scrubbed-140052308.html)…[peace in the Middle East](https://www.bloomberg.com/news/articles/2026-07-17/fears-mount-of-us-and-iran-escalation-after-sixth-day-of-clashes)…[shipping safety](https://x.com/FirstSquawk/status/2078091378383610205)…[deepfake apps](https://www.wired.com/story/san-francisco-demands-apple-and-google-delete-ai-nudify-apps-from-app-stores/)? **The likelihood of Databricks going public:** Despite its [mammoth scale and history of positive cash flow](https://www.cautiousoptimism.news/databricks-continues-to-soar-as-private-company-but-those-margins-are-starting-to-shrink/), The AI and data giant is working to secure $3 billion more, this time at a $188 billion valuation, [which the Journal reports is 40% higher than its December valuation](https://www.wsj.com/tech/ai/databricks-set-to-hit-188-billion-valuation-with-new-investment-from-coatue-9eda18d6). - This is why no one is going public! Why bother exposing your business to the irrational public when you can just slurp a fresh markup for your own equity every six months at the cost of, what, 1.6% dilution? - I’m willing to sell 1.6% of this newsletter for a mere $3 million, a full thousand-fold discount to Databricks’ equity. I promise that will*CO**also*mire in the private markets until it leaves a full-body imprint in the mud. **Netflix, after earnings: **The streaming giant [missed revenue expectations](https://www.cnbc.com/2026/07/16/netflix-nflx-earnings-q2-2026.html) in the [second quarter](https://s22.q4cdn.com/959853165/files/doc_financials/2026/q2/FINAL-Q2-26-Shareholder-Letter.pdf) ($12.56 billion vs. $12.58 billion anticipated), but beat profit estimates by a penny. Slower-than-expected growth and the streaming giant telling investors it would restrict the information it shares about user activity were received like lead weights on a skydiving run. The company’s shares lost around 12% this morning, lowering its market cap to around $315 billion. It’s incredible how, even after covering tech for so long, companies that aren’t pure-play software businesses always feel small: With a market cap of $315 billion, Netflix is worth just 50% more than what Alphabet lost yesterday. The search and ad giant’s shares fell 4% following [reports ](https://finance.yahoo.com/markets/stocks/articles/why-alphabet-googl-stock-falling-005722662.html)that its Gemini 3.5 Pro model was delayed yet again, carving out $200 billion from its market cap — hardly a story given its $4+ trillion market cap. **Csquare and Standard Nuclear’s IPO prospects: **Standard Nuclear (TRISO fuel) [priced ](https://www.renaissancecapital.com/IPO-Center/News/120497/Small-modular-reactor-fuel-developer-Standard-Nuclear-prices-previously-dow)its IPO at $15 per share, below its original range of $18 to $21 per share. The company’s shares closed their first day of trading even lower, at $12.30. Apparently, investors don’t expect small modular nuclear reactors to use TRISO in huge quantities any time soon. I only know that TRISO is one of several fuels that such reactors can use. Sound off in my inbox if you have a theory here, but Standard Nuclear’s IPO was a disappointment. Csquare also [priced its IPO](https://www.renaissancecapital.com/IPO-Center/News/120493/Data-center-operator-Csquare-prices-IPO-at-$21-below-the-range) and listed this week, selling its equity at $21 per share, below its proposed $23 to $27 per-share price range. The stock closed its first day at $20.67 per share, which wasn’t as bad as these things can go. Csquare is a *colocation*-focused data center company, so it isn’t at the precise epicenter of the AI rush. Perhaps that limited investor interest. **China-USA relations**: Relations between the United States and China are trending lower after POTUS [fingered the CCP for election interference](https://www.wsj.com/world/china/trumps-allegations-of-chinese-interference-in-u-s-elections-stir-tensions-with-beijing-c7dd49d5). While the documents released in conjunction with Trump’s speech appear to be more confirmation of ([previously known](https://www.dni.gov/files/ODNI/documents/assessments/ICA-declass-16MAR21.pdf)) Russian [election meddling](https://x.com/willsommer/status/2077932892471067042), blaming China won’t improve relations. This friction is forcing American companies to straddle a widening chasm. A group of publishers ([Der Spiegel, Le Monde, and The Insider](https://theins.press/en/inv/294635)) recently obtained documents that disclose China’s multi-layered plan to degrade and destroy SpaceX’s cash cow, Starlink. The satellite connectivity service is critical to Ukraine’s military operations in Russia, which is a key Chinese ally. - “Meanwhile, Tesla, Musk’s electric car company, whose largest factory is in Shanghai, is heavily reliant on Chinese state loans and the indulgence of the Chinese Communist Party,” the group writes. - Something’s gotta give, politically or economically, between the two nations. Where we’re heading right now looks bad for business and, more importantly, world peace. **Kimi K3 isn’t that cheap** AI lab Moonshot’s newest model in the Kimi series, [Kimi K3](https://www.kimi.com/blog/kimi-k3), is impressive, scoring great marks on several major benchmarks. It reminds me of its predecessor, Kimi K2.5, which was so strong that Cursor chose it as the base for its Composer series of coding models. And if you read public commentary, you’d think that K3 just mogged the world: - Axios: “ [China’s open-weight Kimi model stuns AI world with frontier-level results](https://www.axios.com/2026/07/16/moonshot-kimi-ai-china-model-openai-anthropic)“ - Fortune: “ [Moonshot’s Kimi K3 pushes Chinese AI into Fable-level territory](https://fortune.com/2026/07/16/moonshots-kimi-k3-pushes-chinese-ai-into-fable-level-territory/)“ - Axios: “ [China just erased America’s AI lead](https://www.axios.com/2026/07/17/china-ai-kimi-k3-open-source-anthropic-opus)“ Did it? Our favorite benchmarking group, Artificial Analysis, [ranks Kimi K3 third](https://artificialanalysis.ai/#intelligence), behind Claude Fable 5 and GPT-5.6 Sol. It’s no small feat, given that it’s managed to outstrip Claude Opus 4.8, Grok 4., GPT-5.6 Terra and Luna, not to mention Z.ai’s GLM-5.2 and the latest Muse Spark model from Meta. Moonshot has certainly created the strongest non-U.S. AI model we’ve ever seen, from a raw capability perspective. What’s more, K3 [took first place on a well-known developer benchmark](https://venturebeat.com/ai/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems). But when we consider Chinese models’ usual advantages over American AI, Kimi K3 is a notable exception from the usual talking points: **You can’t run it locally:** At 2.8 trillion parameters, K3 is massive. You cannot run it at home on Ollama. It’s only useful in a centralized-inference setup, so get ready to rent access to it.**It’s not cheap:** At $3 per million input tokens and $15 per million output tokens, K3 is slightly more expensive than OpenAI’s GPT-5.6 Terra (the middle of its current lineup).**It’s token-hungry:** AA notes that Kimi K3[used 130 million tokens](https://artificialanalysis.ai/models/kimi-k3), far above the average of 63 million, in its evaluations. So it’s priced about the same as GPT-5.6 Terra, but[used 34 million more tokens](https://artificialanalysis.ai/models/gpt-5-6-terra)than its most comparable competitor.**Data concerns are real:** At present, you can only use Kimi K3[through Moonshot-mediated inference](https://openrouter.ai/moonshotai/kimi-k3#providers). If you are worried about Anthropic using your usage data to train its models, well, here we are. To sum, Kimi K3 is a big, capable, impressive and somewhat expensive model. That sounds incredibly like a release from an American AI lab, doesn’t it? Will Kimi K3 make it harder for second-tier labs like Google and SpaceXAI to gain market share? I think so. But GPT-5.6 Sol and Claude Fable 5 are already on the market, and are better. Moonshot agrees. At the end of its blog post announcing Kimi K3, the company said that despite “being a highly competitive model overall, K3 nonetheless exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol.” *That’s no sin,* but it has [not moved the Pareto frontier further](https://arena.ai/leaderboard/text/pareto?q=k3) nor broken new ground on sheer capability. Still, K3 may be aces for providing a new, better base for post-training/fine-tuning once it is released more openly, which Moonshot says it’s working on. So Moonshot may eat into some of our closed-source heavyweights’ market share when it’s eventually put through Tinker or a related system. But that won’t happen immediately, and by the time those gains make their way to market, the leading AI duopoly will have newer, better stuff in the market. And so the circle spins again.