Thursday. Are we at war again? No one knows! Stocks are treating the latest military strikes as transient; oil prices are proving stickier. Regardless, everyone’s second-favorite critical trade constriction (shout-out to the Strait of Malacca) is nowhere near prior shipping rates. So, the blockade is still in force, even if unofficially.
But we’re not here to despair about self-inflicted wounds. We’re here to talk business, which means we need to discuss London, SpaceXAI’s new AI model, the major tech companies reshaping the benchmark conversation, and a decline in AI confidence. To work! — Alex
📈 Trending Up #
Ollama, which I love…profitable startups? …Wonder-ful fundraisings…startup-to-startup acquisitions…European compute…Mercor at $20B? …Australian uranium exports…
London: Regular ** CO **readers know that this publication is bullish on
Mistral, the French AI lab that offers both open and closed models, compute, and a Claude Code/Codex competitor. Not only has the company long focused on cheaper models (good in the post-tokenmaxxing era), but its open offerings demand less lock-in than betting your company on a single, closed-source American AI lab.
In early 2025, Mistral opened an office in London. It’s not the only one betting on the rainy city. American cybersecurity company Rubrik has just announced a $500 million investment in Britain over the next five years, establishing London as its European headquarters. (The major AI labs are also increasing their London footprints.)
that Europe just recorded its “strongest quarter in four years for venture funding,” with some $24 billion invested in the second quarter, “up around a third quarter over quarter and two-thirds higher than the $14.4 billion raised in Q2 2025.”Crunchbase News reports- Where is money flowing the fastest? It may be in the UK! Crunchbase continues: “The United Kingdom widened its venture-funding lead last quarter, as startups based in the country raised $10.4 billion — not far from the peak in 2021 at $10.8 billion.”
DeepMind, of course, has strong London links, but demand from AI companies is so strong in the city that The Times reports that “AI companies have taken 661,100 sq ft of workspace in London” thus far in 2026, up from around 500,000 square feet in all of last year. While American technologists and investors tend to view the world as what they can see, I reckon there’s enough activity across the pond to warrant our regular and continued attention.
SpaceXAI made a good model, but OpenAI looms large: The combination of xAI and **Cursor **is bearing fruit, with the pre-merger partners releasing Grok 4.5 yesterday after both companies contributed. The resulting model ranks below Fable 5 and Opus 4.8 from Anthropic, and GPT-5.5 from OpenAI, but ahead of Sonnet 5, GLM-5.2, and every other AI model in existence. Its comparatively low price of $2 per million input tokens and $6 per million output tokens (doubled for requests that exceed the 200,000-token context window threshold) makes it a viable model in price/performance terms.
Fun question: If Grok 4.5 is good enough to drive demand for material inference, how long until SpaceX cuts off its newly cemented compute partners? Either party can pull a 90-day ripcord at will. Call it an inverse-canary for Grok’s success.
Sadly for the SpaceXAICursor boffins, Grok 4.5 came out after the return of Fable 5 (which is better), and on the same day that OpenAI released a new voice model that appears to have cracked the natural conversation barrier that prior models suffered from, by allowing fluid two-way speech; the new version has yet to roll out to ** CO’s devices, so we can’t report testing results this morning.
Worse, OpenAI’s GPT-5.6 models (Sol, Terra, Luna) are slated for release today, likely pushing Grok 4.5 further down the intelligence charts and possibly challenging its value proposition.
**Update: **As I was editing this morning’s newsletter, Meta dropped Muse Spark 1.1, which is not only an incredible upgrade over its previous 1.0 version but also very cheap at $1.25 per million input tokens and $4.25 per million output tokens. Musk Spark 1.1 appears competitive with Opus 4.8 and GPT-5.5, so it’s probably on par with what Grok 4.5 offers, but at a lower price point. Holy shit. Good one, Alexandr Wang and company.
Custom evals, custom models: There’s more! This week, DoorDash and **Databricks **dropped new evals, including DashBench (model pairs for code review and their relative strengths) and the data lakehouse company’s unnamed benchmark comparing model-harness combinations across price and success rates; OpenAI, Anthropic, and Moonshot’s GLM models perform well, using Codex, Claude Code, and Pi harnesses. It will be interesting to see how results from both companies’ benchmarks change once new models are pushed through their testing channels.
- Please make the benchmarks public and update them weekly. Else, what’s the point?
But custom evals are only the start. While the major AI labs bicker for podium positions, rolling your own model is becoming increasingly viable for startups with enough data and gumption. Prime Intellect, a startup that facilitates custom AI model training and continuous improvement, just announced that “demand” for its services has “scaled to over $100M in annualized revenue” in under a year. Hence why it just raised a $130 million Series A.
Why not just use a mainstream model and be done with it? Well, Prime reckons that companies can create better, faster, and cheaper models if they control how models are trained and improved (“**Ramp **trained a 35B model on Lab to beat Claude Opus at spreadsheet search, 27% faster and far cheaper”).
- The argument here isn’t that you train a model once and save money against its use over time; instead, the idea is that your company’s unique data should provide a compounding edge over general models. So, the gains should only increase.
- This is why everyone is yapping about data today; no proprietary data, no custom model; no custom model, no unique edge over your competitors andthe major AI labs.
Meanwhile, more models rained down from the skies. Cognition, the company behind the ‘Devin’ coding tool and owner of Windsurf after that company’s founders jumped ship to Google, released SWE-1.7, which it reckons is on par with or better than many leading models on developer benchmarks and offers strong performance for its cost. And for flavor, Harvey, the legal AI company, is building out its own AI model team, including expansion into additional market categories.
- If all of tech is just bundling and unbundling, I wonder how long it will be before harness companies that make models become model companies that also provide harnesses? Or perhaps the harness is the main product?
Summing: Major AI labs are fighting just as hard today as they have for years to lead the world in AI performance; out of sight, tools to help companies understand and build their own AI models are improving rapidly.
[📉](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)
Meta’s ethics…Salesforce shares, after a downgrade…home sales…giving Nvidia all your money? …Mag7 P/E ratios…Chinese bridges in Iran…Indian mangoes in Japan…
AI confidence: Tired of all the AI talk? I get it. You are not alone. Market demand for debt to finance ever-larger data center capex is fading. That means the money tap for the* infra* that empowers AI could be dialing back. Here’s Bloomberg:
Amazon’s [latest debt sale] was met with a distinctly
[chilly reception], pulling in only 1.6 times as many orders as the $25 billion of bonds offered — a steep slide from the demand it enjoyed just four months ago and well below this year’s average.
Similarly, Business Insider reports that prices for some tech debt are falling: Investor demand for the latest Amazon bond offering was reportedly light, and spreads on other
[bonds issued by Big Tech]companies have widened. The spread refers to the yield investors demand to be paid over a benchmark rate, like Treasury yields.
Not that the tech’s spending binge is slowing down. Meta intends to ramp up its capex spend next year, and the memory crunch is driving bonkers-level investment in new capacity. On one hand, hard-edged spending exuberance from the companies serving AI demand, on the other hand, concern from the folks doing the lending.
You can watch Oracle’s AI-focused bonds here if you want. The company’s 2036 notes carry a 3.85% coupon, but as they are trading at less than par, the effective yield is 6.3%. That’s not great!
And the risks are sky-high. Here’s **Apollo **economist Torsten Slok, from a recent note on AI concern (emphasis original):
The bottom line is that AI has been the one thing holding up both the economy and markets,and with so much riding on so few names, a slower payoff wouldn’t just be a sector problem, it would risk tipping the economy into recession and the S&P 500 into a correction.
Great.