> daily_signal(2026_06_28)

Google capped how much of its Gemini AI even Meta can buy, while a child-safety bill in Congress would quietly build a database of who is behind which account.

PickBits Daily Signal · Sunday, June 28, 2026

By Mark Pickering · 8 min read · June 28, 2026

// tl;dr

Google told Meta, around March, that it could not sell it all the Gemini computing power Meta wanted to buy. The shortfall delayed some of Meta's own AI projects and pushed Meta to tell employees to optimize their token usage. Other Google customers were squeezed too; Meta hit it hardest, and the cap is still in place in late June.
The KIDS Act, which strongly pushes age verification, could reach a US House vote as soon as this week, and you can't verify age online without verifying identity. The mandate would force platforms or third-party vendors to collect identity data, creating a pool the government can subpoena to unmask anonymous accounts, the channel reporters and whistleblowers rely on. Age-check vendors have already been breached.
Coinbase cut its AI spending in half, while usage kept climbing, by routing each request to the cheapest capable model, leaning on Chinese open models GLM 5.2 and Kimi 2.7. CEO Brian Armstrong says better caching took its hit rate from 5% to 60%. Lindy switched to DeepSeek v4 and Snowflake is testing Chinese models, as a price war reportedly brews between OpenAI and Anthropic.
RAISE US launched on June 25: a bipartisan nonprofit targeting $1 billion, over half secured, to retrain US workers displaced by AI. Former Commerce Secretary Gina Raimondo is CEO; founding backers include rivals Amazon, Microsoft, Anthropic, and the OpenAI Foundation. First pilots run in Arkansas, Connecticut, Maryland, and Utah, judged on jobs, not enrollment.

For a year, the running thread here has been that AI's real ceiling isn't how smart the models get, it's supply: whether there are enough chips, and enough power to run them. This week that ceiling pressed on the two companies you'd assume were immune. Google told Meta it couldn't sell it all the Gemini compute it wanted, and Coinbase went the other direction, halving its AI bill by routing work to the cheap Chinese open models we've watched undercut the American labs since May. The same pressure runs underneath the other two stories. A child-safety bill reaches a House vote that, by the simple mechanics of age-checking, would assemble the kind of identity database we've watched Britain build at its border. And the companies automating jobs away put a billion dollars toward retraining the people they displace. The four aren't obviously connected, but each one comes down to somebody upstream deciding how much AI you get, and what you pay for it.

Today: Google capped Meta's Gemini compute, the KIDS Act's age check would create a subpoena-able identity database, Coinbase halved its AI bill on Chinese open models, and RAISE US launched a $1 billion employer-funded program to retrain workers displaced by AI.

1. Google capped how much of its Gemini AI even Meta can buy.

The limit on AI stops being how smart it is and becomes how much you can actually get.

Around March 2026, Google told Meta it could not supply the full amount of Gemini computing capacity Meta wanted to buy. The shortfall disrupted and delayed some of Meta's internal AI projects, and Meta responded by directing employees to optimize their use of tokens, the units that meter how much compute a model burns, and by leaning harder on its own Muse Spark model for critical safety work to cut its reliance on Google. Several other Google clients were affected as well, but Meta felt it most acutely given the sheer size of its demand, and the restrictions remain in place as of late June.

For a year, the throughline in this newsletter has been that compute, not cleverness, is the binding constraint on AI: data centers take years to build, the power to run them is contested in city after city, and the biggest buyers pre-commit capacity long in advance. What is new is the rationing reaching a customer like Meta. And this isn't Google playing favorites or punishing anyone. It genuinely can't sell what its biggest customer wants to buy, which is the scarier version. Even Google is scrambling for capacity: Sundar Pichai said its own compute backlog nearly doubled quarter over quarter, and Google has agreed to pay SpaceX roughly $920 million a month for outside computing power. When the company with its own vast data centers and effectively unlimited cash still gets told there isn't enough, the bottleneck is real, and everyone smaller is downstream of it.

CNBC report June 28 2026 Google limited Meta's use of its Gemini AI models according to a Financial Times report Google told Meta around March it could not supply the full Gemini compute capacity Meta wanted to buy delaying some Meta AI projects Meta told employees to optimize token usage other Google clients affected restrictions remain in place late June — cnbc.com · June 28, 2026

Why this matters: If your product runs on a single AI provider, you just watched the biggest buyer in the world get told there wasn't enough to sell. Most teams quietly assume capacity is a checkbook problem: hit a limit, spend more, it goes away. That just failed for a company with a far bigger checkbook than yours. That doesn't mean a shortage hits you tomorrow, but it does mean a written number matters more than a sales promise. Action this week: Find the one model your product genuinely can't run without, and stand up a tested fallback on a second provider before you need it, so a cap or an outage is an inconvenience and not a stoppage. If a vendor has promised you capacity, get the commitment in writing with real numbers, not a handshake. And turn on token-usage tracking now and compare cost-per-task across providers, because Meta's first move under the cap was exactly that, and it's the cheapest insurance you can buy.

cnbc.com: Google limits Meta's use of its Gemini AI models, FT reports (June 28, 2026)
tbsnews.net: Google limits Meta's use of its Gemini AI models: FT (June 2026)
benzinga.com: Google limits Meta's access to Gemini AI models amid rising demand (June 2026)
bloomberg.com: Google caps Meta's use of Gemini AI, FT reports (June 28, 2026)
ground.news: coverage and bias spread — Google limits Meta's use of its Gemini AI models

2. The age-verification bill heading for a House vote would build a database that could unmask reporters' sources.

You can't check a child's age online without collecting an adult's ID.

The KIDS Act, online child-protection legislation that strongly incentivizes or outright requires age verification, could reach a vote in the US Congress as soon as the week of June 28. There is no way to reliably verify someone's age online without verifying who they are, and that is the whole problem: the mandate would force platforms, or the third-party vendors they hire, to collect identity data and tie it to accounts. That creates a new pool of records the government can demand when it wants to learn who is behind an anonymous account — the same channel journalists use to protect a source and whistleblowers use to come forward. And this is not a hypothetical exposure: age-verification vendors have already been breached, spilling exactly this kind of sensitive identity data.

This is the age-check fight we've tracked moving across the internet, from Britain scanning faces at its border to app stores checking IDs, arriving now as US legislation with a real vote behind it. The objection here is narrow, and worth saying plainly. The goal, shielding children online, isn't the problem, and there are less invasive ways to get there: Snap, among others, has argued for age signals checked on the device or operating system, which never pool everyone's identity in one subpoena-able database. The KIDS Act's approach pools it. And once that record exists, it lives by whatever rules come later, not the ones the bill's authors had in mind.

The Intercept June 28 2026 the KIDS Act online child protection legislation requiring age verification could reach a US congressional vote the week of June 28 there is no way to reliably verify age without verifying identity forcing platforms or third party vendors to collect identity data creating a pool the government can subpoena to unmask anonymous accounts the channel journalists and whistleblowers rely on age verification vendors already breached — theintercept.com · June 28, 2026

Why this matters: If you have ever depended on an account that isn't tied to your legal name, to report a story or flag something inside your own employer, this bill is about the database that would link the two. "Verify your age" and "verify your identity" are not separate asks; technically they are the same collection, and a record built to keep kids off a platform is a record that can later answer a very different question about an adult. Action this week: If anonymous speech matters to you, the vote could come this week, so call your representative now and say plainly that you want age checks done without a centralized identity database. If you run a tip line, a SecureDrop-style intake, or anything that depends on pseudonymous accounts to reach sources, re-model your threat surface around the new exposure the mandate creates, not the law's stated intent — the data pool is the risk, whoever ends up holding it.

theintercept.com: The age-verification bill that could unmask whistleblowers and journalists' sources (June 28, 2026)

3. Coinbase halved its AI bill by routing to Chinese open models, and it isn't the only one.

The model doing your company's work, chosen request by request on price.

Coinbase CEO Brian Armstrong says the company cut its AI spending in half even as token usage kept climbing, by adding an automatic router that picks the cheapest capable model for each request, leaning on the Chinese open models GLM 5.2 and Kimi 2.7. A big part of the saving was unglamorous: better caching pushed its cache hit rate from 5% to 60%, and 91% of its developers never exceeded their prior usage limits. Coinbase isn't alone: the startup Lindy switched to DeepSeek v4, and Snowflake is testing Chinese models as alternatives to OpenAI and Anthropic, as a price war reportedly brews between the two American labs.

Since May, we've tracked the setup for exactly this: DeepSeek making steep discounts permanent, its output running a fraction of the American labs' prices, and Microsoft testing a fine-tuned DeepSeek inside Copilot. What was missing was a name willing to show the savings in production, and Coinbase is it. It's not a free swap, though. A cheaper model behaves differently, and "cheapest capable" is carrying a lot of weight in how Coinbase set this up, which is why the router and the caching matter more than which flag is on the model. But the price gap is now big enough that not testing it is its own decision, and one a lot of finance teams are about to start asking about.

The Decoder June 2026 Coinbase joins the rush to Chinese AI models as Western labs face a pricing stress test CEO Brian Armstrong says Coinbase cut AI spending in half while token usage climbed using an automatic router that picks the cheapest capable model per request leaning on Chinese open models GLM 5.2 and Kimi 2.7 caching hit rate rose from 5 to 60 percent 91 percent of developers never exceeded limits Lindy switched to DeepSeek v4 Snowflake testing Chinese models price war between OpenAI and Anthropic — the-decoder.com · June 2026

Why this matters: If your AI bill is climbing, a company you've heard of just cut its own in half without cutting usage, and the main lever was boring engineering, not a magic model. The takeaway isn't "switch to a Chinese model"; it's that you should know which model is doing your work and what each request costs, instead of letting a default decide for you. Action this week: Before your next budget cycle, pilot a price-and-capability router across providers — including open-weight models — and do the caching work first, because that's the cheapest win and it's where most of Coinbase's saving came from. If you sell to or invest in the big model labs, watch this as the real test of the prices that justify their valuations: track whether customers stay and keep spending, not the headline benchmark scores, because the benchmark gap and the bill are now moving in opposite directions.

the-decoder.com: Coinbase joins the rush to Chinese AI models as Western labs face a pricing stress test (June 2026)

4. The companies automating the jobs put up $1 billion to retrain the workers.

The rivals driving the displacement are co-funding the response to it.

On June 25, 2026, RAISE US launched: a bipartisan nonprofit, with former Commerce Secretary Gina Raimondo as CEO and former Indiana Governor Eric Holcomb as co-chair, targeting $1 billion in multi-year commitments, more than half already secured, to retrain and redeploy US workers displaced by AI. The founding backers include direct rivals Amazon, Microsoft, Anthropic, and the OpenAI Foundation, alongside Bank of America, IBM, Cisco, and Eli Lilly. First pilots run in Arkansas, Connecticut, Maryland, and Utah, and the program says it will measure success by whether people land jobs, not by how many enroll.

All year, we've covered the other side of this ledger: layoffs that quietly cite automation, the Stanford research that flagged an entry-level hiring crunch early, executives who spent 2025 warning about job losses and 2026 walking it back. What's new is that the companies driving that disruption are now putting real money behind a soft landing. The cynical read writes itself, the firms automating your role buying themselves some goodwill, and it's worth keeping. But a billion dollars and a deliberately hard yardstick, jobs rather than enrollment, is more than anyone else has actually committed, so it's more useful to watch whether it actually works than to dismiss it on day one. The four pilot states are where that gets decided.

Rockefeller Foundation June 25 2026 RAISE US launches bipartisan nonprofit former Commerce Secretary Gina Raimondo CEO former Indiana Governor Eric Holcomb co-chair targeting 1 billion dollars more than half secured to retrain and redeploy US workers displaced by AI founding backers include rivals Amazon Microsoft Anthropic and the OpenAI Foundation plus Bank of America IBM Cisco Eli Lilly first pilots in Arkansas Connecticut Maryland Utah measured by jobs not enrollment — rockefellerfoundation.org · June 25, 2026

Why this matters: If AI is reshaping your line of work, the companies building it just funded an on-ramp toward the next job and named the first four states it runs in. Whether this is real help or reputation management is exactly what the jobs-not-enrollment metric is designed to expose, which makes it more honest than most corporate programs and worth holding to that standard. Action this week: Look up whether RAISE US is opening in your state and which roles its pilots target, and bookmark the launch details at rockefellerfoundation.org so you catch the next wave when it scales. If you lead hiring or learning-and-development, study the "redeploy, don't lay off" incentive design now — it's both a template you can copy and a funding pool you may be able to tap as it expands beyond Arkansas, Connecticut, Maryland, and Utah.

rockefellerfoundation.org: RAISE US launches, uniting leading employers and bipartisan governors behind American workers (June 25, 2026)
thenextweb.com: AI giants are funding a $1B program to retrain workers they're displacing (June 2026)
axios.com: Anthropic and rivals back a labor-market response to the AI jobs crisis (June 25, 2026)

» What to watch this week

Whether the KIDS Act gets a floor vote, and whether any amendment swaps the identity-collection mandate for a privacy-preserving device-level age check. The fight isn't the bill's goal; it's whether the mechanism has to pool everyone's identity to reach it.
Whether Google's Gemini cap on Meta loosens or spreads to its other large customers. If the restriction holds or widens, compute has moved from a pricing question to a rationing one, and capacity, not capability, is setting the roadmap.
Which other named Western companies follow Coinbase onto Chinese open models, and whether OpenAI or Anthropic cut prices in response. The tell is customer retention at the American labs, the thing that actually underwrites their valuations.
Whether RAISE US closes the rest of its $1 billion and names the specific roles its first pilots retrain people into. A pledge measured by jobs only means something once there are jobs to point to in Arkansas, Connecticut, Maryland, and Utah.

Tomorrow's signal lands here.