ByteDance Profits Plunge 70%: The Cost of AI Arms Race

When I saw the news that ByteDance’s 2025 net profit fell over 70%, my first thought was: the cost of the AI arms race is finally becoming visible.

According to Yicai News, ByteDance’s full-year 2025 net profit dropped more than 70% year-over-year. Domestic revenue grew about 20%, overseas revenue surged over 40%—business is still growing, but profits have been essentially “eaten up.”

Where did the money go? The answer is obvious: AI.

Compute Investment is “the Ticket to AI’s Endgame”

That’s ByteDance’s internal line. In plain English: to survive the AI era, you must spend money recklessly right now.

I did some rough math. ByteDance’s Doubao LLM, from training to inference, requires astronomical GPU resources. Industry estimates put training costs for a trillion-parameter model at tens to hundreds of millions of dollars. And that’s just training—the real money pit is inference. Hundreds of millions of daily user calls mean continuous compute consumption.

ByteDance isn’t alone. OpenAI still isn’t profitable. Anthropic reportedly loses billions annually. The entire global AI industry faces the same growing pains: bright technological prospects, but broken business models today.

Does “Burn Money for Market Share, Monetize Later” Still Work?

The classic internet playbook: subsidize users to gain market share, monetize once dominance is established. This logic faces challenges in the AI era.

First, monetization is hard. AI products have less user stickiness than social products. Users can write copy with ChatGPT, then switch to Claude or domestic alternatives like Doubao or Wenxin Yiyan anytime. Low differentiation, low switching costs, hard to build moats.

Second, cost structures differ. Internet-era marginal costs approached zero—one more user barely added cost. But AI has positive, high marginal costs. Every token generation requires real compute. More users means higher costs.

This creates a paradox: user growth might accelerate losses.

ByteDance’s Dilemma Reflects China’s AI Industry

It’s not just ByteDance. Alibaba, Baidu, Tencent—all face similar struggles.

Alibaba’s Tongyi Qianwen, Baidu’s Wenxin Yiyan, Tencent’s Hunyuan—all represent billion-dollar investments with no end in sight. Model improvements require continuous investment. Today’s leading model might be surpassed in six months.

More troubling, much of this spending is “defensive.” Even knowing short-term profitability is impossible, they must do it. Because not doing it means complete elimination from future competition.

This reminds me of the cloud computing race. AWS lost money for over a decade before profitability. Alibaba Cloud’s path was similarly long. But the difference is cloud’s business model was relatively clear—sell compute, sell services, predictable costs. AI’s business model is still being explored. Nobody knows the final profit path.

How Long Will Investor Patience Last?

ByteDance isn’t public, so it avoids capital markets’ short-term pressure. But for listed AI companies, sustained losses are unsustainable.

In 2025 we saw adjustments: OpenAI pushed expensive enterprise tiers. Anthropic sought more monetization paths. Domestic companies tried similar moves: Doubao launched paid APIs, Wenxin Yiyan offers memberships, Tongyi Qianwen explores B2B partnerships.

But these revenues are drops in the bucket compared to massive investments.

My Take

As an AI industry observer, ByteDance’s profit plunge doesn’t surprise me. It’s inevitable during technology’s explosive growth phase—investment always far exceeds output.

But I worry: how long can this “cost-be-damned investment” continue? If profitability remains elusive in 3-5 years, will capital enthusiasm wane?

For ByteDance, the good news: sufficient cash reserves and financing capacity to sustain longer losses. Its core business (Douyin/TikTok) remains strong, providing “ammunition” for AI investment.

The bad news: AI investment is a bottomless pit. Today’s trillion-parameter models might be replaced by ten-trillion-parameter ones tomorrow. This arms race has no finish line, only escalating bets.

Final thought: while enjoying free AI tools, we might not realize the cost behind them. Every free conversation consumes real compute resources. This industry is making a massive bet that AI will eventually change the world, and they’ll survive to see that day.

ByteDance was just the first to publish its “report card.” More are coming.