Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week via email here.
A draft report leaked from the Treasury Department warns of the risks posed by the hot-burning AI economy, likening it to the dotcom bubble of the early 2000s and noting its circular financial nature that has many of the key players investing in each other. The document, obtained by the nonprofit newsroom NOTUS, reveals more jitters over AI within the Trump administration, which won the support of tech industry elites by promising to ease the way for the proliferation of artificial intelligence into American life and business.
Meanwhile, the stock prices of AI industry players, especially AI chip makers, are sagging this week as investors worry over the real size and progress of the AI boom. Some of those investors may be taking a new look at the larger business proposition offered by the AI industry. It’s all centered on a small group of AI labs in California spending hundreds of billions of dollars to build and then rent huge, opaque, general-purpose large language models.
There’s no doubt that scaling up transformer models to huge sizes has yielded some stunning capabilities, many of them unexpected and not fully explained. Chain-of-thought reasoning, tool use, and long-horizon planning are all examples of such “emergent” behaviors. But developing huge, generalist models is fabulously expensive. Some of the premier U.S. labs, including OpenAI and Anthropic, are doing it on borrowed money, and are subsidizing the per-token rental cost that enterprises and consumers pay to use the models.
At the same time, enterprises are increasingly using smaller open-source or open-weight models for business tasks that simply don’t require the full intelligence of expensive general-purpose frontier models; in fact, the vast majority of business processes can be managed by smaller, specialist AI models.
Even for high-level agentic applications, enterprises are increasingly using far less expensive open-weight frontier models from Chinese labs like DeepSeek, Alibaba, and ByteDance, which are approaching the U.S. frontier models in intelligence and performance. Some enterprises are developing open-source models and infrastructure to use as backup or concurrent systems—especially as the Trump administration has demonstrated a willingness to cut off access to such models at the drop of a hat.
Open-weight models appear to directly attack the competitive “moat” around the U.S. AI labs. “A cheaper product arrives, looking inferior, priced low, good enough for the customers a [market] leader overlooks . . . then it climbs,” is how Howard Yu, professor at IMD Business School in Switzerland, describes the competition. “The incumbent, protecting its own margins, keeps retreating toward the high end until it runs out of room.”
So why are the U.S. AI labs so singularly focused on developing these monolithic frontier models instead of spending more time on smaller, more cost-efficient models with broader market fit? The labs believe that scaling up large language models (and teaching them to code well enough to create new AI models) is the path to AI that’s generally as smart as human beings (AGI). They also believe that the lab—and its home country—that achieves that level of artificial intelligence first wins all the marbles. At that point most digital work will quickly transition to AI systems, the narrative goes, and the labs and their investors will reap massive returns.
But it’s also these large, generalist frontier models that present the most dangerous risks. Putting aside the safety, cyber, and national security risks, the financial risk is considerable. As the Treasury analysts point out in their draft report, today’s AI companies are more deeply entrenched in the U.S. economy than were the dotcom companies. As much as a third of the U.S. stock market is connected in some way to the AI industry, analysts say. So delays or snags in AI’s rollout could affect the whole economy.
In that event, it’s not far-fetched that U.S. AI labs could be declared “too big to fail,” especially if their systems are entrenched in large enterprises and government agencies. The Trump administration has already proposed that the U.S. take an equity share in domestic AI companies.
The AI industry’s bet-it-all-on-red approach may not be the responsible way to usher in what might be the most consequential technology of our lifetimes. If for no other reason than that humans are in the loop and will need time to adapt, our AI transformation is likely to be more like a marathon than a sprint, as the former Facebook security researcher Georg Zoeller puts it in a recent blog post. He argues that U.S. AI companies should aim for a better balance between revenues and cash burn, between today’s business model and tomorrow’s big breakthrough. (This might create some time for AI companies to catch up on safety research investments, too).
“[I]t’s entirely clear that blowing all your capital/gunpowder early in the race, at minimal efficiency, to make larger and larger explosions to draw more and more attention and the next round of capital, is not a winning strategy,” Zoeller writes.
Meta’s Muse Image generation model poses a threat to ad agencies
Meta has launched its new AI image generator, Muse Image, this week. It’s the first image generation model from Meta Superintelligence Labs, the social giant’s new AI group led by the Scale AI founder Alexandr Wang.
Wang explained on X that the new model analyzes the user’s image request by searching out contextual information on the web. This gives the model a better understanding of what the user wants, and increases the chances of its getting it right on the first try.
The model lets users integrate multiple photos into a creation, and offers a sketch tool with which users can mark up an image with edit requests. Users can download the resulting images and share them directly to a chat, story, or feed, Meta says.
But the tool’s biggest impact may be in advertising. Meta says it’ll make Muse Image available to advertisers through Meta Advantage+, the company’s AI-powered system for advertisers on Facebook and Instagram. Until now, advertisers on Meta’s platforms have had to rely on in-house talent or creative agencies (or, more recently, perhaps their own AI accounts) to create their ads. With Muse Image, they’ll be able to AI-generate their ads. Muse Image, mind you, doesn’t do creative—it generates only what the user tells it—but the cost of making various versions of ads could go down so much that A/B testing them on real social media users becomes faster and cheaper.
In April, Meta Superintelligence Labs released its first frontier model, Muse Spark, which provided an improved brain for the company’s Meta AI assistant. Meta described Muse Image as a companion to that work, positioning the new model as a creative tool built on top of an assistant that already has context on the user. Muse Image will likely roll out to advertisers in the next few weeks. Meta also previewed a new Muse Video model, which is coming later.
The major AI labs have abandoned or weakened earlier safety pledges
The biggest of the AI labs—including Anthropic, OpenAI, Google DeepMind, and Meta—have weakened or abandoned earlier safety pledges stating that they would pause development if specific safety thresholds are reached. The findings are part of the Future of Life Institute’s AI Safety Index, released this week.
The report states that some of the AI companies conditioned their pledge to pause development on whether or not their competitors were honoring similar pledges. The Future of Life Institute panelists described the pattern as a “moving goalpost” that has undermined safety frameworks across the industry.
Despite that, Anthropic, OpenAI, and Google DeepMind still hold the top three positions in the Institute’s yearly safety index. Anthropic earned the highest overall grade, leading five of six domains on the strength of its transparency practices, safety framework, technical research, and governance structure. OpenAI led the Risk Assessment domain, based on a broader evaluation suite and engagement with external testers.
Meta moved up from sixth to fourth place. xAI dropped from fourth to seventh, placing it in the same safety class with China’s DeepSeek and Europe’s Mistral.
Anthropic, OpenAI, Google DeepMind, and Meta had banned military applications of their models, but have all reversed or softened the bans since 2024. All four are now pursuing defense partnerships, joining xAI and Mistral, which had never banned military use.
The panel singled out Anthropic for criticism over what it called the company’s “questionable military engagements,” referencing reports that Anthropic’s AI had been used in connection with the Minab school strike in Iran.
Reviewers also found a gap between public safety messaging and actual conduct at Google DeepMind, OpenAI, and xAI, concluding that stated commitments are an unreliable indicator of safety practice at those companies.
More AI coverage from Fast Company:
- The UN has an AI strategy for everyone except the labs that build it
- Don’t become a meme: How to tell your team you’re going AI-first
- This latest frightening ransomware attack was orchestrated entirely by an LLM
- Our interest in AI slop is hitting a ceiling
Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.
