When AI Workloads Exceed 90 Hours: Code Review Under Pressure

The 90-Hour Code Review Grind

If you think code review is a calm, contemplative exercise, you haven't been inside an AI lab lately. Recent reports from OpenAI, Anthropic, and other top players paint a picture of relentless deadlines and weekend marathons. One former OpenAI engineer said his weeks routinely hit 70 hours, and during product crunches, 90-hour weeks are the norm. That's not a typo. Ninety hours.

What does that mean for code review? When you're exhausted, your judgment slips. You approve patches you'd normally question. You miss the subtle race condition or the off-by-one error that could take down a service. The pressure to ship the next big model doesn't just affect the researchers—it cascades down to every engineer who has to review their work.

Token Loans: A New Way to Fund AI Infrastructure

Meanwhile, in China, banks are getting creative about financing AI companies. The Bank of China recently launched something called a "Token Loan" in Guangzhou's Haizhu district. It's a financial product where the loan amount is tied to how many tokens a company's AI models consume. They've already approved 28 million yuan for five companies, with 8 million actually disbursed.

Why tokens? Traditional loans rely on physical collateral, but AI startups are mostly intangible. Token consumption is a proxy for real business activity—if your model is being called a billion times a day, you're probably doing something right. The bank's chief economist called it a "penetrating insight" into a company's operations. For code review, this means financial pressure to demonstrate token usage, which could lead to inflated metrics or sloppy practices just to hit numbers.

When Your IPO Is Too Hot to Share

Speaking of pressure, Unitree Robotics just went public with the lowest lottery rate in STAR Market history—0.018%. That's absurdly low. Over 9.78 million investors tried to get shares, and only 19,414 got lucky. One winner told reporters he didn't dare post about it on social media, fearing jealousy from coworkers and requests for loans from relatives.

For engineers at such companies, the IPO windfall can be life-changing, but the scrutiny is intense. Code review becomes even more critical when your company's valuation is tied to every line of code. A single security flaw could tank the stock.

ByteDance's New AI Data Security Division

ByteDance, the parent of TikTok, has created a new division called "AI Data & Security" that sits alongside its Seed and Flow groups. It merges several data teams to provide cross-modal data services for all of ByteDance's large models. The head, Wang Yinglei, is overseeing a massive reorganization that started in June 2026.

From a code review perspective, this is a big deal. Centralizing data security means more oversight, but also more bureaucracy. Reviewers will need to navigate new approval chains and compliance checks. It's a double-edged sword: better security, but slower iteration.

Google DeepMind's Pivot Away from Flagship Models

Google DeepMind is reportedly cutting a third of its staff and stepping back from chasing frontier models. Instead, they're focusing on smaller, more efficient "Flash" models that are cheaper to run. This is a strategic retreat from the model-size arms race, and it has implications for code review.

Smaller models mean less code complexity, but also less margin for error. And with layoffs, the remaining engineers will be stretched thin, reviewing more code with fewer resources. The pressure to maintain quality while cutting headcount is a recipe for burnout.

DeepSeek's Botched Release and Open-Source Harness

DeepSeek had a rough week. They quietly released DeepSeek-V4-Pro-0813, then pulled the announcement less than 24 hours later. The API docs still list the model, but the official notice is gone. Meanwhile, they open-sourced their Harness agent framework under the MIT license, which is a boon for developers who want to build custom AI agents.

For code review, open-source Harness is a goldmine. You can inspect every plugin, every loop, every sandbox. But the botched release raises questions about their internal review process. How did a model with such a short public life get through QA? It's a cautionary tale about rushing releases.

Anthropic's 2028 Revenue Projections and Watermark Plans

Anthropic is reportedly targeting $190-200 billion in revenue by 2028, which would justify a $2 trillion IPO. That's SpaceX-level ambition. They're also acquiring Israeli AI company Decart for $6 billion, their largest acquisition yet.

On the technical side, they're adding invisible text watermarks to Claude outputs to comply with EU regulations. This has sparked debate: will watermarks degrade code review? If you ask Claude to review your code and then copy the suggestions, the watermark might persist, potentially flagging your work as AI-generated even if you wrote it yourself. It's a mess that needs careful thought.

Practical Advice for Reviewing AI-Generated Code

So, what can you do to keep your code review sane in this chaos? First, set hard limits on hours. If you're working 90-hour weeks, you're not reviewing effectively. Second, use automated tools to catch the boring stuff—formatting, syntax, common patterns—so you can focus on logic and design. Third, don't trust AI-generated code blindly. Always run tests, and if the AI wrote it, treat it like a junior developer's code: review it with extra care.

Finally, remember that code review is a human activity. It's about communication, mentorship, and building trust. No amount of AI can replace a thoughtful, well-rested reviewer who asks the right questions. So take a break, sleep, and come back to that pull request with fresh eyes. Your future self—and your users—will thank you.