Google quietly raised its 2026 infrastructure budget last earnings call. The capex guidance went from up to $190 billion to $205 billion. That's $15 billion added in just 90 days.
Meanwhile, if you look at the AI tools you actually use every day, it's the opposite story. Token prices are falling, and OpenAI is reportedly considering price cuts.
Money pouring in like crazy while prices drop? That contradiction is worth digging into.
Here's what's actually going on
Google's Q2 wasn't bad on paper. Revenue hit $119.8 billion (up 24% YoY), and cloud revenue jumped 82% to $24.8 billion, blowing past the $22.4 billion estimate. But the stock still fell more than 7% after earnings.
Capex is the culprit. CFO Anat Ashkenazi explained the increase is "primarily due to an acceleration in the delivery of capacity to meet growing demand". Here's the thing: that number jumped $15 billion in just three months. The Verge read it as the company "admitting it can't accurately forecast its own costs".
And it's not just Google. Microsoft (roughly $190 billion), Amazon (consensus raised to $207.4 billion), and Meta ($138.9-145 billion) are all on similar trajectories this earnings season. Combined 2026 capex across four hyperscalers is projected to hit $695 billion, climbing to $870 billion in 2027. Cowen analyst Derrick Wood warned: "If they raise capex again, based on what we saw in the reaction of Google, it's probably going to lead to selling pressure".
But the numbers get weirder
Normally, "spending more" reads as a good sign. This time it read as the opposite — because capex is growing way faster than revenue.
Sequoia's David Cahn flagged an "approximately $600 billion annual revenue gap" between AI infrastructure spend and actual AI revenue. That gap has only widened. Forbes' analysis puts the current AI capex-to-revenue growth divergence at roughly 46% — already bigger than the 32% divergence seen right before the 2001 telecom bust.
Jefferies went even further, warning that "massive capital destruction" in the US is the most likely long-term outcome as Chinese open-source models erode American hyperscaler profitability. Here's why they think that.
Meanwhile, AI is basically on a fire sale
Chinese open-source models processed 36.4 trillion tokens in the week ending July 19 — up from 4.4 trillion in late April, an 8x jump in three months. US models processed just 7.4 trillion in the same week.
Price is why. According to OpenRouter's Justin Summerville, Chinese open-source models run 60-90% cheaper than leading Anthropic and OpenAI models. Z.AI's GLM 5.2 scores within a percentage point of Anthropic's Opus 4.8 on benchmarks — at roughly a fifth of the cost.
There are real switching stories already. AI automation startup Lindy AI moved 100% of its traffic from Anthropic's Claude to DeepSeek. CEO Flo Crivello said it would "save Lindy millions of dollars within months".
That pressure has reached OpenAI and Anthropic too. Reports say OpenAI is internally reviewing steep token price cuts, and CEO Sam Altman admitted "cost has become a very big issue".
| US frontier models | Chinese open-source models | |
|---|---|---|
| Token price | Baseline | 60-90% cheaper |
| Benchmark | Opus 4.8 baseline | GLM 5.2, within 1pt |
| 3-month token volume growth | - | ~8x |
| Risk | Price rigidity, vendor lock-in | Data governance, support stability |
Put it together: infrastructure players (Google, Microsoft, Meta) are burning cash faster than revenue grows, while service prices (tokens) are getting cut in half by Chinese competition. Both trends can't hold at once. Whichever one breaks first will hit the price tag on the AI tools you use.
Heads up
There's no guarantee this price war continues. If Jefferies' "capital destruction" scenario plays out, the surviving vendors could regain pricing power. Don't mistake today's fire sale for the new normal.
So here's what to actually do
- Audit your vendor dependency
Check if your traffic and workflows are locked into a single AI vendor. 100% dependency means you eat 100% of any price hike. - Reassign models by workload
Keep frontier models for complex reasoning and coding; route summarization and repetitive tasks to cheaper open-source models. You don't need Lindy's full switch. - Be careful with long-term contracts
Don't lock in annual deals just because today's pricing looks good. Keep room to renegotiate quarterly. - Line up an open-source backup path
Get an API key for something like GLM 5.2 even if you're not using it yet. It's leverage when your main vendor raises prices. - Watch this earnings season's capex numbers
If Microsoft, Amazon, and Meta keep raising capex, that's a signal infrastructure cost pressure is headed toward service pricing.
Go deeper
AI's finally expensive enough to make Wall Street nervous The original piece that seeded this article — first to lay out Wall Street's reaction to Google's capex announcement. theverge.com
Alphabet Q2 2026 earnings live updates Real-time coverage of revenue, capex, and stock reaction on earnings day. cnbc.com
The AI Capex-To-Revenue Gap Is Widening An analysis comparing the capex-revenue divergence to the 2001 telecom bubble, with the numbers to back it up. forbes.com
Chinese AI models are gaining ground The most detailed breakdown of Chinese open-source model pricing and token share. cnbc.com
Jefferies' "capital destruction" warning The starkest take on the long-term risk to US Big Tech's AI investment. aninews.in
