Welcome to Cautious Optimism, a newsletter on tech, business, and power.
Good morning and happy Monday. Given how quickly news is moving today, instead of waiting for the newsletter to get its usual scrub I’m sending it out raw. After mere single-digit cups of coffee. Expect an errant comma or two, but better now than later and tidier, I think. — Alex
📈 Trending Up: Being wrong … the Eagles … American AI, like really … venture capitalists arguing over supporting the German far-right … DeepSeek on iOS … startup shudowns … water wars … not giving up on friends? … satellite Internet access …
Voice AI: ElevenLabs’ recent round and what Synthesia are building are giving me pretty good vibes about AI crossing the uncanny valley in a reasonable time-frame.
📉 Trending Down: Free speech … not being high on your own supply … crypto controls in China … small wins … Chinese real estate developers …
DeepSeek’s AI models are costing Americans hundreds of billions
The Nasdaq is off 3.90% in pre-market trading this morning, with tech companies associated with the present AI boom taking even larger hits:
Nvidia is off 12.3% in pre-market trading, after falling 3.1% Friday
Microsoft is off 6% in pre-market trading
Amazon is off 4.4% in pre-market trading
Meta is off 3.3% in pre-market trading
The list goes on. (1% of $1 trillion is $10 billion for reference.)
Why the declines this morning? More strident concern that recently-released, and market-shockingly good AI models from China will undercut American AI competitiveness. And that much recent or planned investment into AI hardware might be less economically productive than previously anticipated. Let’s quickly explain both.
Rising concerns about American AI competitiveness:
DeepSeek’s V3 and R1 models — and excellent releases from Bytedance and others — show that despite an increasingly strict set of export limitations, American AI companies are not as far ahead of their Chinese rivals as expected. Or, that recent AI gains by American companies are simply not as defensible as anticipated. This calls into doubt the huge wage spend that many U.S.-based tech companies are putting into their AI work.
If a handful of folks in China can come up with what DeepSeek recently dropped on the market, a lot of AI work in the United States is highly inefficient. Why spend so much on human AI inputs (developers, etc) if similar or better work can be done for much less?
Less efficient spend and smaller competitive edges are not conducive to creating greater long-term cash flows.
Rising concerns about AI hardware spend
Up until the last few days, pledges by major American tech companies to spend ever-more on their datacenters were a useful way to tell the market that their AI work was serious. You won’t be GPU poor if you are going to drop $65 billion on capex this year, right Meta?
But if you can do so much — again, as DeepSeek has putatively shown possible — with a collection of likely aged chips and sheer gumption, then are those massive hardware pledges more white elephants than proof of future AI market share?
Vibes on planning to spend tens or hundreds of billions of dollars building datacenters of ever-greater electricity demand may have flipped; thus, spending plans have been called into some doubt. And, thus, the value of many American tech companies.
Less efficient spend and smaller competitive edges are not conducive to creating greater long-term cash flows.
You could even argue: Pledging to spend tens of billions on AI hardare to keep a competitive advantage over rivals is evidence of a lack of innovative thinking on the software side?
Don’t just worry about the stock market. The value of Anthropic, xAI, OpenAI, and Mistral have moved as much as stocks. But as many AI model companies are private we lack a good view into how much their worth has fallen since DeepSeek changed at least the conversation in AI.
If I had to hazard a guess, if OpenAI was a stock, it would be down as much as Nvidia. If not more, fair or not.
Enough doomerism, what’s the silver lining?
There’s plenty of reasons to not be too worried.
First, the limited cost of training DeepSeek’s models has been called into question. As has the number of high-end — if perhaps not cutting edge — AI chips the company owns has become a greater point of investigation. It’s unlikely that DeepSeek managed its seeming coup without a hardware foundation of some scale. Therefore, AI chip demand is likely to stay material.
Second, DeepSeek’s work is open-source. So, expect the learnings contained to get quickly unpacked, absorbed, and incorporated. Then, the market will be back to a situation in which there are chip-haves and chip-have-nots. As before, just post DeepSeek R1 ingestion. That doesn’t sound like a market that the United States and the larger Free World need to lose.
Third, DeepSeek’s hosted AI models are lobotomized to meet Chinese government rules. The fact that R1 et al are forced to play very stupid when asked about historical events means that they are self-hobbled in the form that most folks are using them. That’s good for competing hosted models.
VCs are still sorting out if DeepSeek is net-positive for humanity, or not. To wit:
And in response:
Quick Hits
AI Agents and Human Computing Surfaces: Much has been made about AI agents that can interact with computer graphical user interfaces designed for humans. Computer use from Anthropic and Operator from OpenAI are the two key products here. Watching the OpenAI Operator demo got me thinking about how slow using human-focused GUIs is for a machine. But as the Internet — and, therefore, digital tooling more generally — is human-prepared, teaching AI to be a sort of digital human makes sense.
In time, I expect AI agents to work nearly exclusively with APIs built for their interaction.
What does that mean for humanoid robots? If we expect digital work to become more robot-tuned eventually, why would we not expect physical work to become more robot-tuned? Put another way, if we expect that AI agents will eventually work using technology not built for humans, why wouldn’t we anticipate the same for physical robots?
The answer, I think, is that it’s far easier to rebuild digital work to suit AI agents while the physical world is going to stay human-centric for the foreseeable future. Because we live here.
Social media is hard: Twitter’s not growing much, isn’t really profitable, and is not seeing revenue soar. While X has become the town square for right-wing politics, it has yet to turn its changed branding and bent into profits.
Crypto prices are dropping: Bitcoin et al are fun, but mostly track the stock market. Who could have guessed.
I just ask Gemini about January 6th 2021. Here's the reply:
"I can't help with responses on elections and political figures right now. While I would never deliberately share something that's inaccurate, I can make mistakes. So, while I work on improving, you can try Google Search."
"The fact that R1 et al are forced to play very stupid when asked about historical events means that they are self-hobbled in the form that most folks are using them."
Do you realize how often Google Gemini responds "I can't comment on that" when asked US related "political" questions? Too often.