Today, we're looking at tech valuations, the latest AI models from OpenAI and SpaceXAI, a little startup deal that made us smile, and the return of Microsoft's AI game. #
Wednesday. I’m hosting a venture capital roundtable for TWiST today, so check your podcast feeds this evening for the latest.
Today, we’re looking at tech valuations, the latest AI models from OpenAI and SpaceXAI, a little startup deal that made us smile, and the return of Microsoft’s AI game. To work! — Alex
📈 Trending Up #
Apple and Broadcom, domestically…free speech in Florida, via the courts…investing in chips…small EVs in the United States! …going public, buying other companies…hype about the value of data (again)
Inflation: Hopes that the Iran War could be resolved through diplomacy have been dashed. The ceasefire failed to halt fighting in Lebanon, Iran shot at a few tankers, the United States struck Iran, and now we’re back to where we were.
In response, stocks are falling, and oil prices are rising.
- If you want tech companies to thrive, you want low interest rates and a hot public market. But rising energy prices will not improve the chances of an interest rate cut.
- The Iran War’s intensity is going to be inversely correlated to the frequency of tech exits.
Blue Origin’s valuation: CNBC reports that space company Blue Origin is raising $10 billion, its “first outside funding round that will value the rocket company at $130 billion.”
Jeff Bezos is said to be kicking in $2 billion, and Coatue is said to be investing $4 billion, with the rest coming from institutional investors. Similar to how Alphabet likes to raise external capital for Waymo, it makes good sense to allow others to buy into Blue Origin.
- Blue Origin recently had a major setback when its flagship rocket explodedduring a test, harming the company’s infrastructure and setting its timelines back. - But if we’ve learned anything from how investors value SpaceX, lots of folks expect the space industry to scale into something truly celestial.
The sun shines on OpenAI: The AI lab said late last night that its GPT-5.6 Sol (large), Terra (medium), and Luna (small) models will “launch publicly” this Thursday. Axios reports that OpenAI had to work with the Trump administration to get out of the default penalty box, and received permission to release its work after “additional testing [by the Center for AI Standards and Innovation] and meetings between the company and government officials.”
AI model reviews from early testers usually lean positive, but the GPT-5.6 Sol hype train is off the rails: “truly amazing,” “incredible,” “an execution beast,” “a very, very good model,” “excellent judgment,” “an amazing model,” and so on on.
From the back of the classroom: Tomorrow we’ll also see SpaceXAI releasing Grok 4.5, which Elon Musk describes as “an Opus-class model, but faster, more token-efficient and lower cost.” That slots this modelbehindOpenAI and Anthropic, but it may be on par or even slightly ahead of open-weight AI models from China.Meta’s moves: Besides working on an improved Muse Spark model, Meta Superintelligence Labs has built and released an image-generating model called, you guessed it,Muse Image. The model will help powerMeta’s first-party apps.
Yes, but: Over in China, MiniMax is said to be putting the finishing touches on a new, 2.7-trillion-parameter model (Grok 4.5 is a 1.5-trillion-parameter model, for reference).
MiniMax’s models are very popular, as they somehow combine near-frontier intelligence with low cost. If the new model proves impressive, it could chip away at the pricing power of American AI models. But if China puts the kibosh on new, powerful open model releases from its domestic labs, we’ll never know.
Speaking of which: After DeepSeek raised $7 billion, Z.ai is planning to raise around $4 billion by sharing more of its stock post-IPO.
Buying startups? On the heels of Figma buying a vibe-coding startup, unicorn Vercel (Next.js) is now acquiring Better Auth.
The gist, per The New Stack, is that Better Auth (raised $5 million) was moving into the agentic identity market, which will help bolster Vercel products like Connect and Eve. Such deals are critical to the health of the startup sector. More, please!
Tech earnings: Dutch chip tooling firm ASML reports next Wednesday, and TSMC,** Netflix**, and Wipro will share their performance a day later. The following week, will see Tesla, IBM, ServiceNow, Alphabet, Intel, and SAP reporting.
As always, we’ll be looking out for how AI impacts these tech businesses, their spending plans, and what they’re seeing in the way of demand.
[📉](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)
Personal privacy…remote work…cheap smartphones…building integrity in NYC…Google’s lead in image generation
Being too small to go public: AI creative unicorn Higgsfield AI has reached annual run-rate revenue of $500 million, joining ElevenLabs.
Reaching run-rate of $100 million is passé now; $500 million may be the new Big Boy startup demarcation line.
Microsoft is taking control of its AI future #
Every AI month is equal to half a year of the traditional business cycle. Every AI quarter is a year. Every AI year is a decade. Compounding is one hell of a drug.
Microsoft once reigned the global AI game. Its early investments and partnership with OpenAI gave it both ownership and access to the world’s leading AI IP, and a massive cloud customer to boot.
Today, the company has invested in Anthropic, loosened its ties with OpenAI, and somehow found itself trailing its hyperscaler rivals when it comes to first-party AI products.
At least that was the case. In early June, Microsoft unveiled a staggering amount of AI products. More than a half-dozen new AI models, a project that would make Windows agentic, new hardware, a personal agent, a few frameworks, and a few other things (Microsoft Scout is still in preview, sadly).
If you were hoping to see Microsoft act like a real AI lab with an attached enterprise and consumer software business, its 2026 Build releases were precisely what you ordered: near-term upgrades, new products in the works, and tinkering for the years ahead. Redmond’s new AI models include a trillion-parameter model (MAI-Thinking-1), and a quick coding model (MAI-Code-1-Flash). The company stressed their price-to-performance ratio, with MAI-Thinking-1 claiming a “smaller inference footprint than much larger models,” and MAI-Code-1-Flash can supposedly solve “harder problems with up to 60% fewer tokens.”
The models are also entirely homegrown:
MAI-Code-1-Flash: “We trained it from the ground up on clean, traceable and enterprise-grade data, without distillation from third-party models.”MAI-Thinking-1: ” We trained it from the ground up on clean, traceable and enterprise-grade data, without distillation from third-party models”
Can’t fault a company for reusing a turn of phrase that it likes!
So what do we have in hand? Microsoft is building models it considers comparable to middle-tier offerings from labs like Anthropic, with a focus on low cost — quick improvements are also promised, but we can’t vet that claim just yet.
What is the company going to do with those models? Here’s a hint from a few weeks ago:
Axios:Microsoft weighs DeepSeek for Copilot Cowork Lowering inference costs for its products seems to be the impetus behind that move. But why use a Chinese open-weight model when your own in-house model is ready for use? Microsoft apparently reached the same conclusion, and is moving to test its own AI models inside its enterprise software stack.
Here’s TechCrunch from earlier this week: Indeed, when it comes to two of its most widely used programs — Excel and Word — Microsoft has begun to use its homemade
[MAI models]to respond to a certain percentage of user prompts, Bloomberg[reported]Tuesday. In the past, the company had advertised the fact that large parts of Office 365 are[powered by models from both OpenAI and Anthropic].
Clearly, the company’s still testing its use of MAI models inside critical customer work surfaces. But if those models perform as well as Microsoft hopes in production, it could shed third-party models and stop funding OpenAI and Anthropic’s margins.
I don’t use Microsoft 365, so I can’t comment on how AI+Office is shaking out in practice (can’t wait for its Q2 results!). When the company made Copilot Cowork available in June after a few months of testing, it said “more than half of the Fortune 500” was already using the product, and it is “the fastest growing feature in the history of our Frontier program, and Cowork has among the highest user satisfaction of any Copilot or agent experience we have shipped.”
We’re probably talking about a lot of expenditure here.
- MAI-Code-1-Flash could also make it cheaper to use GitHub and its agentic harness, similar to how Cursor turned to its Composer model series to cut inference costs.
It’s too early to declare Microsoft a first-party AI leader à la Anthropic and OpenAI. Still, the company is taking the required steps to join that rarefied pantheon, alongside Alphabet and (perhaps) Meta. By this time next year, I presume the MAI models will be one option amongst several offered to Microsoft’s enterprise customers.
Not that we have any indication of this happening, but: If Microsoft wanted to take the AI world by storm, it could release its models on an open-weight basis. There’s probably less revenue to be made from such a strategy, but there’s a lot of potential market share to capture, provided China really closes the door on its own labs’ open releases.