How did Nvidia find itself in a lose/lose game?
And: I am This Week in Startups’ token liberal
Welcome to Cautious Optimism, a newsletter on tech, business and power.
📈 Trending Up: NYC fintech … ‘a Russian nesting doll of bullshit’ … OpenAI’s flop era … yes … AI lobbying … stablecoins … bitcoin outflows … SLMs … spine … streaming music prices … sure, Jan …
📉 Trending Down: House majorities … rule of law … small business … Alibaba profitability … Nvidia’s hegemony? … German GDP growth …
Things That Matter
Genesis Mission: A national effort to put “Federal scientific datasets” to work, training “scientific foundation models and create AI agents to test new hypotheses, automate research workflows, and accelerate scientific breakthroughs” is a good idea. Building an “American Science and Security Platform” that will “serve as the infrastructure for the Mission” by providing “high-performance computing resources,” “AI modeling and analysis frameworks,” and “secure access to appropriate datasets, including proprietary, federally curated, and open scientific datasets” is a good idea.
You can read the entire EO here, but working collectively at the Federal level to provide data and foment scientific discovery in the AI realm makes good sense. To wit:
Perhaps the above helps explain Amazon’s new, $50 billion push to build compute for the government? It also adds context to the October-era announcement from the Department of Energy to build new supercomputers with HPE and AMD.
Free trade is losing: The NYTimes wrote up the Federal government’s push to take stakes in private companies, a regular theme here at CO. It’s a good summary if you want one, but I want to flag the White House’s comment on the piece for our consumption:
“If business-as-usual policies worked, America would not be reliant on foreign countries for critical minerals, semiconductors and other products that are key for our national and economic security. The administration’s targeted equity stakes ensure that taxpayers get a good bargain and that the ball meaningfully moves forward to encourage further investment by the private sector.”
Put another way: If the free market worked, then the United States would have retained market share in less economically-productive sectors, leading to a glorious world in which the nation was poorer. Therefore, we shall ape the Chinese government’s industrial policymaking, and if you complain about it, you hate America.
It’s weird how much phrasing matters!
I’ve joked that I am This Week in Startups’ token liberal, but an issue where I’ve often found myself more sympathetic to the Republican party than the Democrats over the years is trade. It’s great. Big good. Very nice. Efficient. Productive. Powerful.
Reading missives from the modern right, however, makes it clear that POTUS’s tariff regime is not merely being tolerated by many political conservatives but embraced. And not solely because it grants the current head of the Republican Party more power; because the strain of conservatism that values Small Town Living has its back up that those same hamlets have become economic backwaters.
The shift is also deeper than broadsides from Senator Hawley on the WTO. A recent essay in The Federalist argues that:
The unrestrained worship of markets has created a widening chasm between the few and the many, thanks to conservatism’s embrace of globalization and mass labor importation. Championing GDP maximization above all else led to globalization and the offshoring of American jobs. This has not only decimated entire regions like the Rust Belt and hollowed out communities due to the loss of manufacturing but also led to a feeling of despair that has fueled an opioid epidemic, all while enriching America’s chief geopolitical rival: China. […]
We are told this is the price we must pay for our system of “free market” capitalism, but there are two problems with this that conservatives must grapple with.
The same piece goes on to argue that “capitalism […] is not sacrosanct.” Oh?
You could toss the VP’s vituperation against tech companies’ use of H1-B visas in the same bucket, frankly. But trade. Trade seems to have become a bad word as the world grapples a rising re-bloc-ificaton, with China, Russia and friends in one corner, and the United States, Europe, and other democracies on the other. Just as during the first Cold War, economic might will win. And we’re going to need lots more trade to win the upcoming challenge. We will not win by becoming Diet Coke China.
I wonder what part the chest-thumping America doesn’t need allies rhetoric from the White House adds to progress made in making comparative advantage feel dirty.
Just kidding: Yesterday, CO argued that Google had pulled ahead of the AI market thanks to the launch of Gemini 3 Pro. The model is gosh-darn impressive, and, along with a host of other AI model releases last week, helped underscore that there’s still large intelligence gains to be found in how we build LLMs.
Enter Anthropic! The AI lab dropped Claude Opus 4.5 yesterday — family Claude genus Opus species 4.5, if you will — its latest large flagship model. And it’s hot damn, golly-gee good. And perhaps best of all, while still not cheap Opus 4.5 will cost:
Opus 4.1 (old. worse): $15 per million input tokens, $75 per million output tokens
Opus 4.5 (new, better): $5 per million input tokens, $25 per million output tokens
The new model is still expensive compared to GPT 5.1 and Gemini 3, but oh so much cheaper than its predecessor. Good!
How did Nvidia find itself in a lose/lose public opinion scrap?
What’s a GPU worth over time? The answer to that question is worth hundreds of billions of dollars of corporate profitability and trillions of dollars of collective technology market cap.
If GPUs from Nvidia and its rivals burn out quickly from sustained AI usage, major tech companies may be underreporting hardware depreciation costs, allowing them to report better-than-reality net income.
If GPUs from Nvidia and its rivals don’t burn out quickly from sustained AI usage, major tech companies may be able to wring enough revenue (and gross profit) from their lifespans to render their depreciation calendars accurate, and their capex defensible.
The shouting over just how quickly GPUs go the way of the buffalo once plugged into a data center and pounded with AI training and inference is getting nasty. Investor Michael Burry of Big Short fame managed to kick up enough dust that Nvidia sent a memo to investors arguing that his points are specious, his math spurious, and his conclusions a sham. (Barron’s broke the news.)
Here’s the company’s partial response (only part of the memo is publicly available) to the GPU depreciation argument: