Enterprises were willing to shrug off data concerns when AI models didn't cost so much. But now that AI's expensive, why not build your own? #
Monday. Stocks are down, oil prices are up, and the Strait of Hormuz is still a mess.
After the tempo of air strikes picked up, the United States reinstated its blockade of Iranian ships through the waterway, and is saying it will charge tolls. It’s Groundhog Day, again. If you were hoping for lower interest rates predicated on falling inflation rates thanks to declining fuel costs, this is bad, bad news.
Today, we’re looking at Helsing’s massive funding round, Meta’s ever-rising capex commitments, Nvidia as an AI model developer, and the enterprise data mess. To work! — Alex
📈 Trending Up #
Charging tolls in the Strait of Hormuz…threats to speech…AI job loss fears…antitrust…Intel in Ireland…fast-fashion IPOs…regulatory capture…nepotism…leaving venture for AI(really!) …
Helsing: German defense unicorn Helsing raised a $1.8 billion Series E, taking its valuation to $18 billion. The company last raised €600 million Series D last June. The company said it remains majority European-owned and that demand for equity in its latest sale outstripped the shares offered.
Like Anduril, Helsing fuses AI and warfighting hardware into a single platform, a product that modern militaries are loving as the speed of warfare spikes and evolves to include more drones, favoring computer speed over muddy human boots.
- There’s massive demand for companies building autonomous warfighting systems in Europe. German startup Quantum Systems raised a $1.2 billion Series D in Juneto expand its fleet of intelligent hardware (airborne drones today, with sea and land coming). - There are enough European, AI-infused, defense tech startups coming up that the “European Anduril” mantle is being
hotly contested. Meta’s compute footprint: Reports that Meta would consider renting some of its servers to other AI companies were read as a tacit admission that the company had overspent on GPUs, racks, and HVAC.
So much for that. After releasing Muse Spark 1.1 and offering it on a managed basis (using its own compute), the company is expanding a Louisiana-based data center campus, adding $40 billion to the project, which will be worth “more than $50 billion.”
It’s also working with local power companies to offer “approximately $2.65 billion in total customer benefits,” meaning that local residents may benefit from the data center project.
That’s the right move: lowering public discontent about data centers is a key wrench in any intelligent AI company’s toolbox.
[📉](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)
Chip stocks…teenage social media access…data center economics…not working for Anthropic…getting along…open AI in the United States? …smartphone sales…
Is America losing the open-weight AI race? While it’s fun to track demand for the latest AI models from the likes of OpenAI and SpaceXAI, it’s worth keeping a closer eye on Nvidia.
The chipmaker’s Nemotron 3 Ultra model is currently generating nearly as much inference on OpenRouter as all xAI and OpenAI models combined. This is because Nemotron 3 Ultra is free on the platform at present — in exchange for letting Nvidia train its models on user data — and it’s a fast and capable model that’s affordable on a per-task basis (AA data).
It’s also incredibly open-source. You can download its “pre-trained, post-trained, and quantized checkpoints along with the datasets used for training,” which is super cool.
Now, Nemotron 3 Ultra is not on par with leading Chinese models when it comes to capability, but by making it free, the chip giant is receiving immense amounts of usage data which it will use to improve subsequent iterations. You don’t eat such costs if you don’t have big plans.
If Nvidia’s next Nemotron models can close the intelligence gap with closed-source models while keeping their open-source soul, then we may already have the West’s response to Chinese open-weight models.
- Z.ai’s founder is making the casethat China’s potential move to wall in domestic, open AI models would be a mistake. How much weight one entrepreneur’s voice carries isn’t clear, but it’s nice to see that the peoplemakingopen AI in China aren’t excited about further government meddling.
The enterprise strikes back #
As the AI cost crisis begins to fade from the forefront, a related fear is rearing its head: The need to protect enterprise data from AI labs. As with cost, concerns about data are nothing new.
But as the AI industry matures, enterprises are growing apprehensive. The cost of AI usage mattered less when inference loads were modest, and similarly, data concerns could be waved away when enterprises were only planning to use AI, rather than actually doing it.
Several voices in the tech world are proposing one solution to both concerns: Lower-cost, open AI models trained on private data for internal use.
Let me catch you up…
Subscribe to Cautious Optimism to unlock the rest.
Become a paying subscriber of Cautious Optimism to get access to this post and other subscriber-only content.