AI's Hidden Toll: How Assistive Tech Workers Are Burning Out

The AI Grind: 90-Hour Weeks Are Now the Norm

If you've ever wondered what it's like to work at the bleeding edge of AI, the answer might surprise you: it's exhausting. Recent reports from insiders at top AI labs like OpenAI and Anthropic paint a picture of relentless crunch. During what they call “sprint periods”—those intense pushes before a big product launch or model release—some employees are logging 90-hour weeks. That's not a typo. Ninety hours.

One former OpenAI engineer told reporters that his typical week was 70+ hours. He left for a startup, thinking things would calm down. Now he works 50–60 hours—until a launch looms, and then it's back to weekends and late-night firefighting.

For assistive technology, this is more than a labor story. The people building the tools that help disabled users navigate the world are running on fumes. Burnout doesn't just hurt them; it slows down innovation that could change lives.

Token Loans: A New Way to Fund AI Startups

Meanwhile, banks are getting creative. In Guangzhou, China, the Bank of China has started offering something called a “Token Loan.” It's a financial product designed for AI companies that don't have physical assets to use as collateral. Instead of a factory or property, they put up their token consumption—the number of tokens their AI models process—as proof of business activity.

So far, the bank has approved about $3.9 million in loans across five companies, with $1.1 million already disbursed. The idea is that token usage is a real-time measure of how much a company's AI is being used. It's a clever workaround for a fundamental problem: AI startups are often “asset-light,” but they still need capital to buy compute and pay engineers.

For assistive tech startups, this could be a lifeline. Many are small, mission-driven teams with limited funding. A loan based on actual usage—rather than a traditional balance sheet—might let them scale without giving up equity.

Manus Goes Independent: What It Means for Users

In a move that caught many by surprise, Manus—an AI assistant that had been under Meta's wing—announced it's spinning off to become an independent company again. The transition will affect some users' data. If you signed up with an Apple or Facebook account, you'll need to watch for in-app notifications. Data from after December 29, 2025 will be deleted between August 23 and 24, 2026, to comply with local regulations.

Manus says it will store data in the US and Singapore going forward, and there's a “welcome back” reward for users who return. It's a reminder that the AI landscape is shifting underneath us. For assistive tech users who rely on tools like Manus for daily tasks, these transitions can be disruptive—but independence might also mean more focus on user needs.

Nvidia's Stake in SpaceX: A Side Note on AI's Web

In a filing with the SEC, Nvidia revealed it holds a chunk of SpaceX stock—about $17.2 billion worth. That's because Nvidia invested $10 billion in xAI back in January, and when SpaceX acquired xAI in February, the investment converted into SpaceX shares. It's a tangled web, but the takeaway is simple: AI and space are becoming intertwined, and Nvidia is betting big on both.

For the assistive tech world, this matters because Nvidia's chips power many of the AI models we use. Their financial moves signal where the industry is heading—and that could trickle down to the tools available for disabled users.

Anthropic's Watermark: A Double-Edged Sword

Anthropic has started embedding invisible watermarks in text generated by some Claude models. The goal is to help identify AI-written content, which could be useful for accessibility—for instance, verifying that a transcript was auto-generated. But there are concerns: Will watermarks interfere with copyright? What if you just use Claude to polish a sentence? Will that count as AI-generated?

Anthropic says it will offer a free API so anyone can check for watermarks, which is good. But the tech is new, and the implications for assistive tech users—like those who rely on AI to write emails or documents—are still unclear. A watermark shouldn't be a scarlet letter.

Google's DeepMind Cuts: Focusing on Flash, Not Flagships

Google DeepMind is reportedly shifting away from chasing the biggest, baddest models. Instead, they're focusing on their Flash line—smaller, faster, cheaper models that are easier to run. That could be great news for assistive tech, where efficiency often matters more than raw power. If you're building a real-time captioning app, you don't need a model that can write a novel; you need one that can keep up with speech.

But the shift comes with a cost: DeepMind might cut a third of its staff. That's a lot of talent, and it could slow down long-term research. For assistive tech, the hope is that the Flash models get better and cheaper, but we might lose some of the moonshot projects that could have led to breakthroughs.

The Human Cost of AI's Race

All these stories point to a bigger issue: the AI boom is burning out the very people building it. At Meta, employees describe being “conscripted” into AI teams without much choice. At startups, the pressure is just as intense. A study from UC Berkeley found that while AI tools speed up individual tasks, they don't reduce the workload—they just make the pace faster and the tasks pile higher.

For assistive technology, this is a warning sign. The field needs sustained, thoughtful development, not a marathon of 90-hour weeks. Burnout leads to mistakes, and mistakes in assistive tech can have real consequences for people who depend on it.

What's Next: Hope for a Sustainable Path

Despite the gloom, there are reasons to be hopeful. The rise of token-based loans could make funding more accessible. The shift to efficient models like Flash could reduce costs. And the growing awareness of burnout might push companies to rethink their culture.

For those of us who use or build assistive tech, the message is clear: we need to support the humans behind the machines. That means advocating for reasonable work hours, investing in sustainable practices, and celebrating the small wins—like a new feature that makes a screen reader a little faster, or a loan that lets a startup hire another engineer.

The future of AI is exciting, but it should not come at the cost of the people who make it possible. Let's build tools that empower everyone—including the builders.