Google Is Speedrunning Its Own AI Graveyard

Last week, developers sat down to work, opened Google’s AI coding tool, and found it had uninstalled itself overnight. No warning. The editor they used every day was gone, replaced by an empty chat box. To get the old setup back, you had to wipe everything, chat history included.

Now the part that should bother you more. This happened during the best week Google’s AI division has had in years. A faster model, custom chips nobody else can touch, and a planned $190 billion in spending. So how does a company that is clearly winning spend one week making its own paying users hate it? The answer is not the one the headlines are selling.

Google’s best week in years

On paper, Google I/O 2026 was a victory lap. Google walked on stage with Gemini 3.5 Flash, a model fast enough to beat last year’s top-tier flagship at writing code, plus a new agentic coding platform to run it on. Then came the money. CEO Sundar Pichai put Alphabet’s 2026 capital spending at around $190 billion, roughly six times what it spent four years earlier.

That is not the budget of a company losing a race. According to the I/O developer keynote, millions of developers already build on Google’s models every month. And then, in the very same week, Google turned around and throttled the exact people paying for all of it.

2026 capex $190B

Alphabet's planned spend on AI infrastructure this year, roughly 6× the ~$31B it spent in 2022.

New model 3.5 Flash

Gemini 3.5 Flash launched at I/O, fast and beating last year’s flagship at writing code.

Monthly devs 8.5M

Developers building on Google’s AI models every month, by Alphabet’s own count.

Three broken promises in seven days

When you pay for a developer tool, there is an unspoken deal: the thing that worked yesterday still works tomorrow. Google broke that deal three separate times in a single week, and not one of the breaks was about the technology. Each one was about the contract underneath it.

First, the usage limits. Google moved its paid plans onto a compute-based system where every prompt costs an invisible, shifting amount, and hitting the weekly cap locks you out for up to seven days. Second, it gave the Gemini command-line tool an execution date. Third, the self-uninstalling update from the intro. The top comment on Google’s own launch announcement was not about the shiny new model. It was a one-line joke about the limits, and it out-voted the news.

7 days Usage limits

Paid plans moved to compute-based caps. Hit the weekly limit and you are locked out for up to seven days. Even failed generations that returned nothing still drained the budget.

June 18 CLI killed

The Gemini command-line tool many developers had wired into their daily workflow got an execution date and a 28-day window to migrate to a replacement Google admitted was not yet at feature parity.

Wiped Self-uninstall

An Antigravity update overwrote and broke existing installs overnight. Clawing the old setup back meant wiping everything, chat history included.

Why did the Gemini usage limits cause a revolt?

Because the meter was invisible and the punishment was huge. Under the new compute-based caps, paying customers reported burning a five-hour quota in under an hour, with no clear readout of what each prompt actually cost. Worse, even failed generations counted. If the model errored out and handed you nothing, that still drained your budget.

To Google’s credit, it panicked fast. Within about 48 hours it tripled the Antigravity limits, then tripled them again and reset everyone’s quota. That tells you two things. Google does listen, once the noise gets loud enough. And it shipped those original caps without ever testing them against one real day of work.

Is the AI pricing squeeze industry-wide?

Yes, and this is the fair defense Google rarely gets. GitHub Copilot is moving all subscribers to usage-based billing. Cursor reworked its plans the same way. Anthropic added weekly rate limits, and OpenAI raised prices too. The era of cheap, all-you-can-eat AI is ending, and not purely out of greed: autonomous coding agents chew through compute in a way a chatbot never did, and somebody has to cover the bill.

One inconvenient fact cuts the other way. Google was the last of the big three to raise its Flash-tier price. So if everyone is doing this, why did Google’s version land like a scandal? Timing and execution. Google stuffed the bad news inside its biggest keynote of the year, so the celebration and the price hike arrived in the same envelope. While competitors wrote blog posts explaining the why, Google changed things quietly and let developers find out by getting locked out.

• Cursor Reworked plans onto usage-based pricing First mover
• Anthropic Added weekly rate limits to Claude paid plans 2025
• OpenAI Raised prices across its paid tiers 2025
• GitHub Copilot Moving all subscribers to token-based billing June 2026
• Google Raised its Flash-tier price and added compute caps Last to move

Same move across the board. Only one of them let paying users find out by getting locked out.

Is Google losing the AI race?

No, and it is not particularly close. The new model is fast and beats Google’s own previous flagship at code. Google designs its own AI chips, the TPUs, which its biggest American rivals have to rent rather than build. It also hands out capable open-weight models for free, which OpenAI and Anthropic do not.

Then there is the fact that should reframe the entire story. Anthropic, the company furious developers are switching to, runs Claude on Google’s chips. Google is so far ahead on raw infrastructure that its single biggest rival is also a paying customer. A company losing the AI race does not have the competition renting its hardware. So this was never about whether Google can build great AI. It can. The problem is something no spec sheet measures.

The headline "Google lost the AI race"

vs

The receipts Strong model, own chips, free open weights

Fastest-tier model

Gemini 3.5 Flash beats Google’s own previous flagship at code, on day one.

Custom TPUs

Google designs its own AI accelerators. Its US rivals rent compute; they don’t build the silicon.

Free open models

Google ships capable open-weight models for free, which OpenAI and Anthropic do not.

Rival = customer Anthropic, the lab furious developers are switching to, runs Claude on Google’s TPUs. A company losing the race does not have its biggest rival renting its hardware.

The Google graveyard, and the Microsoft twist

That something is trust, and it is a hole Google dug by itself. There is a community-run site called Killed by Google that lists every product Google has launched and then shut down, and the count is past 300. Messaging apps killed one after another. Cloud services with paying customers switched off. And that Gemini CLI from a minute ago? On track to be killed inside of a year.

There is a twist. Microsoft has actually discontinued even more products than Google has. So why does Google wear the graveyard reputation while Microsoft walks free? Style, not volume.

Microsoft tends to give long runways and clean upgrade paths. Google kills fast, kills with barely any notice, kills things mid-pivot, and renames the survivors so often you lose track of what you are running. So when Google throttles a plan or force-migrates a tool, developers picture one thing: the next headstone. The reputation turns every small stumble into proof of a pattern. That is the cost of 300 dead products: your next mistake gets charged at ten times the rate.

Killed by Google 300+

Products launched and then shut down, per the community-run Killed by Google list.

Gemini CLI < 1 yr

Time from launch to a posted kill date for the CLI, the newest headstone in the row.

Reputation tax 10×

How much louder each new stumble lands once a company is known for abandoning what it ships.

Why does Google keep killing its own tools?

The cause is structural, and it has a name developers use only half-jokingly: promotion-driven development. Inside Google, the rewards flow to the people who launch new things, not to the people who keep old things alive, boring, and working. Maintenance does not write a promotion packet. A shiny launch does.

So you get the pile. Antigravity, the Antigravity CLI, Jules, AI Studio, Code Assist, Firebase Studio, all overlapping, all half-competing. Google does not lack direction. It has about eleven of them. Will it kill its shiny new coding platform next year? Probably not, since it has bet the developer story on it. But it will likely rename, re-price, and fold it into something else twice before then. Welcome to the permanent reorg, which costs a business built on top about as much as a funeral.

~11 overlapping, half-competing AI dev tools, each one somebody’s promotion packet

AntigravityAntigravity CLIGemini CLIJulesAI StudioCode AssistFirebase Studio …and counting

Rewarded Launch a new thing

Invisible Keep an old thing alive

What should developers actually do?

Treat platform risk as a line item, not a vibe. When you pick a Google AI tool, assume it can be re-priced or renamed inside a year, keep your workflow portable, and avoid wiring anything load-bearing into a tool that is one keynote old. None of that means avoiding Google. The models are good and the chips are unmatched. It means pricing in the churn before it prices in you.

And watch one signal over the next month: whether the next model, Gemini 3.5 Pro, ships on time and ships clean. If it does, last week was a bad week. If it does not, last week was a trailer. Google does not need a smarter model to fix this. It already has one. It needs the harder, more boring thing: pick its tools, keep them, and stop ambushing the people who pay for them.