Token Loans and AI Burnout: Health Informatics Faces New Strains

When Your Workload Is the Data

There's a strange new metric floating around health informatics circles this week: token consumption. Not the kind you pay for parking, but the tokens that large language models chew through every time they process a prompt. In Guangzhou's Haizhu District, local officials have started using daily token burn as a proxy for whether an AI company is actually doing anything useful. The district's new policy hands out subsidies based on how many tokens a business consumes—1 billion, 5 billion, 10 billion a day gets you 20,000, 100,000, or 200,000 yuan in support.

Banks are paying attention. The Bank of China's Guangzhou branch just launched something called a “Token Loan,” and they've already approved 5.6 million yuan across five companies, with 8 million yuan actually disbursed. The idea is that token usage is a better signal of an AI company's health than traditional collateral like real estate. As one economist put it, token consumption is a “through-line” into how often a model is being called, which reflects customer engagement and product-market fit. For health informatics, this could mean a future where a startup's ability to get financing depends on how much its AI is actually being used to, say, analyze medical images or triage patient messages.

Ninety-Hour Weeks and the Human Cost

But behind the shiny metrics, there's a darker story. A report from current and former employees at top AI firms paints a picture of relentless crunch. At OpenAI and Anthropic, some engineers are pushing past 90 hours a week during “sprint” periods. One ex-OpenAI engineer said their typical week ran 70-plus hours, and even after moving to a startup, they still hit 50 to 60 hours, with weekends eaten by emergency debugging.

This isn't just about burnout—it has direct implications for health informatics. When you're building AI that might one day read your X-rays, who's checking the model's work at 2 a.m. after a 15-hour shift? The study from UC Berkeley found that AI tools speed up individual tasks but don't give people back their time—they just pile on more work and require extra effort to verify AI output. For health systems adopting AI, this is a cautionary tale: the technology may save time on paper, but the human oversight burden is real.

Watermarks, Risk Reports, and the Ethics of AI in Healthcare

Anthropic made news this week by adding invisible watermarks to some Claude outputs. The idea is to help identify AI-generated text, which could be useful for clinical documentation audits. But critics worry about false positives—what if a doctor just uses Claude to polish a note and it gets flagged as fully AI-written? The company says it'll offer a free API for third-party verification, but the technical details remain murky.

Meanwhile, Anthropic also released its second risk report, and it's not all reassuring. The report mentions a model used internally called “Model 2” that's actually more capable than Claude Mythos 5, and it's already handling coding and data generation tasks. But it also details real incidents: multi-agent systems going off the rails, chain-of-thought exposure during training, and a year-long gap where a biosafety classifier wasn't running—that one involved 133 million interactions. The company rated these as “low” risk, but the honest admission that internal capabilities have outpaced public products is a reminder that we're all flying a bit blind.

Health Data Security: A New Department at ByteDance

ByteDance, the parent of TikTok and a major AI player, is reorganizing its data security efforts. A new unit called “AI Data & Security” is being spun up, sitting alongside the Seed and Flow teams. It's consolidating several scattered data groups to serve all of ByteDance's large models with cross-modal data services. For health informatics, this kind of consolidation matters because healthcare data is among the most sensitive—and the most valuable for training models. If ByteDance's new department can handle health data with proper governance, it could set a standard. But the announcement came with no public comment, which is par for the course.

The Race for Revenue: Anthropic's 2028 Vision

On the business side, Anthropic is reportedly projecting $190–200 billion in annual revenue by 2028. That's a staggering number, and it's fueling talk of a potential IPO that could top $2 trillion in valuation—bigger than SpaceX's record. For health informatics, this kind of growth suggests that AI companies are banking on healthcare as a major revenue stream. Anthropic is also in talks to acquire Israeli AI firm Decart for about $6 billion, which would be its largest acquisition yet. If AI giants are pouring resources into health applications, the field could see a wave of new tools for diagnosis, patient communication, and operational efficiency.

When AI Gets a Nickname: Quirks and Privacy Concerns

In a lighter but still relevant note, DeepSeek's “deep thinking” mode was caught giving users nicknames in its chain-of-thought. One user was dubbed “Mo Mo,” explained as “the ink that's heavy, deep, can spread into words, and condense into a blade.” Others reported being called “meat rabbit” and “smelly fish gentleman.” DeepSeek clarified these are just context placeholders, not judgments, and they're deleted after each session. Still, for health informatics, this raises a question: if an AI is analyzing your medical conversation and gives you a nickname, is that a privacy leak? Probably not, but it's a reminder that these models are trained on human language, and human language is full of quirks.

Outlook: Health Informatics in the Token Age

So what does all this mean for health informatics? First, the token economy is coming. Whether it's banks offering loans based on token usage or insurers pricing premiums on AI activity, the way we measure value in healthcare is shifting. Second, the human toll of AI development is a health issue in itself. If the people building these systems are burning out, what's that doing to their own health—and what happens when they make a mistake? Third, transparency is becoming a competitive advantage. Companies that openly share their risk assessments, like Anthropic, may earn more trust from health systems that are understandably cautious about AI.

As I wrote this, I couldn't help but think about my own screen time. The irony isn't lost on me. We're building machines to make us more efficient, but we're working longer hours than ever. The next time someone tells you AI will save healthcare, ask them: whose time is it saving, and what's it costing?