Maxxing Out
The Credit Strategist - July 2026
AMZN’s management was pretty upset to discover a $500 million monthly bill for token usage in May as employees engaged in massive tokenmaxxing. The company immediately placed caps on token usage, but the damage was done. It wasn’t just bloated usage but the high cost-per-token that delivered a wake-up call to managers at AMZN and elsewhere whose employees were gaming the system to try to impress them. And now, as The Wall Street Journal reports (“OpenAI and Anthropic Are Facing A Price War,” June 13-14, 2026), “Big companies and startups, chafing at rapidly escalating artificial intelligence costs, are increasingly turning to tools that tap into cheaper AI models, including some from China.” The cost-saving alternatives include switching among a mixture of third-party AI models and in-house systems based on free open-source models. But the big news is that real pressure seems to be building on the most powerful (and most expensive) AI models sold by OpenAI and Anthropic as cheaper alternatives become available.
Vishal Misra, vice dean of computing and AI at Columbia University’s engineering school, notes that “you don’t need a model that knows quantum gravity. These open-source models are very capable, and the ability to charge a big premium for AI is going to diminish.” MSFT unveiled smaller AI models that operate more efficiently than leading-edge models while NVDA launched Nemotron, a family of cheaper models. Open-source Chinese models from Alibaba and DeepSeek and lesser-known companies like GLM also offer cheaper alternatives as do Small Vertical Language Models (SLMs) that can perform more modest tasks and run on desktop computers or even mobile devices.
The argument that Chinese models are inferior to those built at much higher cost by Western companies is coming under question. Part of that cost was driven by salaries that would make even profligate NBA owners (who overpay for mediocrity as one of their basic business practices) blush. Michael Green made this argument in his Substack column: “By optimizing training stability rather than relying on brute-force compute infrastructure, labs like DeepSeek and tech giants like Xiaomi have triggered a permanent, deflationary price war.” (“A Token China Shock,” June 28, 2026) He points out that DeepSeek’s R1 and Xiaomi’s MiMo-V2.5 models are undercutting major Western competitors by 90-99%, dropping the cost of tokens to fractions of a cent per million. This doesn’t mean that these Chinese models can do everything that Western models can do. But they don’t need equivalent capabilities to provide what users require. As noted above, you don’t need a supercomputer to solve a simple algebra problem.
Token prices will be a key issue in these companies’ forthcoming IPOs. These pressures are leading OpenAI and Anthropic to consider sharply lowering their prices which would likely slow their revenue growth even if usage numbers keep rising. OpenAI is now reportedly considering delaying its IPO until next year to secure a trillion-dollar valuation which Sam Altman seems to view as some kind of dick-measuring test, but this could backfire if pricing pressures continue or intensify as is likely. He also reportedly offered the U.S. government a 5% ownership interest in the company. Further price pressures will only hurt the valuations of both OpenAI and Anthropic. True believers won’t be deterred, however. Highly respected tech investor Gavin Baker said on the June 26th All-In Podcast that he believes Anthropic would be worth $3 trillion as a public company mentioning the price war. With users doing everything possible to lower their token costs, the remarkably high margins cited by Mr. Baker in touting such a high valuation won’t be sustainable, however. He may need to rethink. A lot of AI usage will be required to lift Anthropic’s valuation to such an exorbitant level.
Intensifying price competition on top of opaque and increasingly leveraged financing structures also poses potential financial stability risks. The circular nature of the AI industry’s financing arrangements coupled with a reported $1.8 trillion of off-balance sheet obligations raises serious questions about what could happen if the ambitious financial projections of the AI hyperscalers and chip manufacturers don’t materialize. At the very least, the gigantic market caps of some of these companies may come under even more pressure than they’ve seen recently.
The Bank of International Settlements (BIS) warned about this in its latest Annual Report:
“The opacity of AI-sector financing compounds these vulnerabilities. Hyperscalers, chip makers and AI labs are linked through a complex web of private arrangements. The most prominent is circular financing: chip makers and hyperscalers take equity stakes in AI labs or neocloud providers, who in turn commit to multi-year purchases of chips or computing power. Data centre construction is increasingly outsourced to third parties that lease facilities back to hyperscalers on long-dated contracts with embedded exit clauses. The terms of such deals are typically poorly disclosed, with risks of the same asset being pledged multiple times. Together, such arrangements account for a sizeable share of sector-wide financing and forward revenue.”
The BIS also points to potential risks in the opaque private credit sector: “Any tightening in credit conditions could expose existing vulnerabilities in the less transparent private sector space, whose reach has expanded among middle market and small firms.” (25) As indicated by use of the adjective “existing” in that sentence, credit conditions in private credit are already tightening as lenders pull back on lending to software and related companies while dealing with higher levels of stress in existing portfolios and demands for redemptions for significant percentages of their funds.
Finally, META’s announcement that it plans to build out a new cloud business to help recoup its enormous spending on AI infrastructure raises questions about long-term AI demand. This news follows moves by SPCX to lease large amounts of xAI’s excess data center capacity to Anthropic and GOOG. If these hyperscalers are suddenly looking to offload capacity, what does that say about their outlook for AI demand? Is this just a short-term measure to lower costs or a signal about long-term demand trends? It certainly suggests that data showing that only a small percentage of announced capacity additions are currently under construction (less than 5%) or fully funded (less than an additional 18%), suggesting much of that future capacity may never be built or will be delayed significantly. I note that META’s announcement sent CRWV stock down sharply; CRWV is a company I’ve pointed out is particularly vulnerable to a slowdown in AI demand due to its lack of competitive advantages, history of losses, and heavily leveraged balance sheet. As noted above by the BIS and elsewhere (including in this newsletter), if AI demand doesn’t materialize as quickly or as robustly as its promoters expect, there are going to be serious losses among equity and debt investors. CRWV and APLD are at the top of that list.
