Google Starts Rationing Meta's Access to Gemini as the Compute Crunch Hits the Cloud

TECH BRIEF

Monday from Front Research, and the two pressure points of the AI trade, who makes the memory and who controls the compute, both moved on the same day. In Seoul, Samsung and SK Hynix are set to commit as much as 2,000 trillion won, roughly $1.3 trillion, to new chip capacity over the next decade. In Silicon Valley, the Financial Times reported that Google has started rationing Meta's access to its Gemini models because it cannot supply all the compute Meta wants to buy. Washington cleared Anthropic's Mythos 5 model for about 100 companies and agencies, with the more capable Fable 5 expected to follow. And the month ends with the Magnificent Seven down roughly $2 trillion in market value in June, the group's worst stretch in years, as investors keep asking when the spending pays off.

• Samsung and SK Hynix are expected to unveil up to 2,000 trillion won ($1.3 trillion) of investment over 10 years at South Korea's presidential office, with as much as 900 trillion won aimed at a new southwest semiconductor cluster and each company building four to five fabs

• Google has limited Meta's purchase of Gemini model capacity after telling Meta around March it could not meet demand, delaying some of Meta's internal AI projects and pushing Meta to tell staff to ration AI tokens, according to the Financial Times

• The US Commerce Department cleared Anthropic to release its Mythos 5 model to roughly 100 vetted companies and federal agencies on Friday, and talks are advancing to restore access to the more capable Fable 5

• The Magnificent Seven shed about $2 trillion in market value in June, more than two-thirds of the S&P 500's monthly decline, with Microsoft on track for its worst month since December 2000 as doubts grow over the payoff from more than $700 billion in 2026 AI capex

• Why it matters: Supply, compute, regulation, and price all moved against the AI trade at once, and the market's June verdict shows investors are no longer taking the payoff on faith.

Here we go.

Samsung and SK Hynix Line Up $1.3 Trillion in Chip Spending as the Memory Race Goes National

• Samsung Group and SK Group are poised to announce as much as 2,000 trillion won, about $1.3 trillion, of investment over the next decade as part of President Lee Jae Myung's industrial strategy

• The plans were set to be presented at a briefing at the presidential office in Seoul on Monday afternoon, with up to roughly 900 trillion won earmarked for a new semiconductor cluster in the country's southwest

• Samsung Electronics and SK Hynix are each expected to build four to five fabs in the Gwangju and Honam area, with Samsung adding chip packaging plants in South Chungcheong and SK Hynix expanding NAND capacity in North Chungcheong

• The commitment lands after a brutal week for both stocks, which each fell roughly 9% to 12% during the global memory selloff

• Why it matters: The two companies that supply most of the world's AI memory are locking in a decade of capacity, turning the current shortage into a state-backed, structural buildout.

South Korea's two largest conglomerates are preparing the biggest industrial commitment in the country's history. According to reporting from Seoul, Samsung Group and SK Group are set to pledge up to 2,000 trillion won, roughly $1.3 trillion, over the next ten years, with the centerpiece a new semiconductor cluster in the southwestern Honam region. Samsung Electronics and SK Hynix are each expected to build four to five fabrication plants in and around Gwangju, while Samsung adds advanced packaging capacity in South Chungcheong and SK Hynix expands its NAND output in North Chungcheong. The package was due to be unveiled at a 2 p.m. briefing at the presidential office on Monday, presented as the flagship of President Lee Jae Myung's growth agenda. (Fortune, Bloomberg, Korea JoongAng Daily)

The timing is striking. The announcement comes days after both stocks were hammered in the worst memory selloff in years, each sliding roughly 9% to 12% as investors worried that AI memory demand had run ahead of itself. By committing to a decade of fab construction backed by Seoul, the two companies are betting the opposite, that high-bandwidth memory scarcity is durable enough to justify hundreds of billions in new capacity. Reports put the southwest cluster alone at up to 900 trillion won, with the balance of the headline figure spread across packaging, NAND, and related infrastructure. (Seoul Economic Daily, Bloomberg)

Why it matters: Samsung and SK Hynix together control the lion's share of the DRAM and high-bandwidth memory that every AI accelerator depends on, so a coordinated, government-blessed $1.3 trillion buildout is one of the most important supply signals in the entire AI complex. For investors, it cuts two ways. The bullish read is that the suppliers closest to the data are putting their balance sheets behind a long memory super-cycle, which supports pricing and the broader infrastructure thesis from Micron to the equipment makers. The bearish read is the one the market has been pricing all week: capacity this large, brought online over a decade, is exactly what eventually breaks a shortage and turns record memory margins back into the brutal commodity cycle these names are famous for. The fact that it takes a national industrial policy and a presidential ceremony to underwrite the spend also tells you how capital-intensive and strategically fraught AI memory has become. The number to watch is not the headline trillion, it is the timing of when this supply actually ships, because that is the moment the current pricing power is most at risk.

Google Starts Rationing Meta's Access to Gemini as the Compute Crunch Hits the Cloud

• The Financial Times reported that Google has limited Meta's purchase of Gemini model capacity after Meta sought more than Google could provide

• Google told Meta around March that it could not meet the full capacity Meta wanted to buy, and the shortfall disrupted and delayed some of Meta's internal AI projects

• Other Google clients have been affected to a lesser degree, but Meta's unusually large demand made it the most constrained

• In response, Meta has told staff to be more efficient with AI tokens, the units that measure model usage

• Why it matters: When even Google cannot sell a customer as much compute as it wants, the AI bottleneck has moved from chips to the cloud capacity that runs them.

Google has started capping how much of its Gemini models Meta can buy, according to a Financial Times report published Sunday. The constraint dates to around March, when Google told Meta it could not supply all the Gemini capacity Meta had sought to purchase, a shortfall that the report says disrupted and delayed several of Meta's internal AI initiatives. Google sells access to Gemini to outside companies through its cloud, and Meta had become one of the largest buyers of that capacity even as it builds its own models and, as covered last week, its own accelerators. (CNBC, Yahoo Finance, Business Standard)

The internal response is the most telling detail. Meta has reportedly urged employees to use AI tokens more efficiently, a striking message from a company spending well over $100 billion this year on its own infrastructure. Other Google customers have felt the squeeze too, the report says, though none as acutely as Meta, whose demand outstripped what Google was willing or able to sell. The episode lands in the middle of a week dominated by exactly this theme, that the scarce resource in AI is no longer just leading-edge silicon but the data center capacity, memory, and power needed to put that silicon to work. (CNBC, Business Standard)

Why it matters: This is a small story with a large implication. For Google, rationing a marquee customer is a sign of strength, evidence that Gemini demand exceeds even its enormous capacity and that its cloud is supply-constrained rather than demand-constrained, which is the better problem to have. For Meta, being told to conserve tokens while spending record sums on its own buildout exposes how far behind demand the whole industry's compute remains, and why Meta is racing to bring in-house accelerators online. The broader read-through for investors is that the AI bottleneck is migrating up the stack, from GPU allocation to the cloud capacity, memory, and power that turn chips into usable intelligence, which is precisely why the hyperscaler capex numbers keep climbing. It also complicates the tidy picture of frontier labs renting compute on demand; if the largest, best-capitalized buyers cannot get what they want, smaller AI companies have even less leverage. Watch whether this pushes more of the majors toward self-supply, because every Meta that hits a capacity wall at a rival's cloud is a stronger argument for owning the whole stack.

Washington Clears Anthropic's Mythos 5 for Wider Use, With Fable 5 Next in Line

• The US Commerce Department gave Anthropic permission on Friday to release its Mythos 5 model to roughly 100 vetted companies and federal agencies

• The clearance follows a roughly two-week period in which the government barred Anthropic from giving foreign nationals access to Mythos 5 and Fable 5, prompting the company to disable all global access to both

• Anthropic co-founder Tom Brown and other executives met with Commerce Secretary Howard Lutnick and senior officials to negotiate the resolution

• The Commerce letter did not lift restrictions on the more capable Fable 5, but reporting indicates Anthropic is close to winning that access back as well

• Why it matters: The first big AI restriction of this cycle is being unwound deal by deal, showing the government now sets the terms on which frontier models reach the market.

The standoff between Anthropic and Washington took a major step toward resolution on Friday, when the Commerce Department cleared the company to release its Mythos 5 model to about 100 approved companies and federal agencies. The clearance reverses part of an order from roughly two weeks earlier that had barred Anthropic from giving foreign nationals access to Mythos 5 and Fable 5 over fears the models' safety guardrails could be circumvented, a restriction so broad that Anthropic responded by switching off global access to both systems entirely. (Fortune, TechCrunch, CNBC)

Getting here took direct negotiation. Anthropic co-founder Tom Brown and other senior executives met with Commerce Secretary Howard Lutnick and administration officials in recent days to hammer out the terms. The Friday letter pointedly did not address Fable 5, Anthropic's most capable model, but the company said it expects talks to continue and aims to restore access soon, and separate reporting describes Anthropic as close to winning back both Fable 5 and full Mythos 5 availability. (Fortune, Stocktwits)

Why it matters: This is the other side of the same coin as last week's news that Washington asked OpenAI to stagger GPT-5.6's release. Taken together, the two episodes show an administration that is now actively gating who can use the most capable American models and on what timeline, not through legislation but through case-by-case approvals negotiated in cabinet offices. For Anthropic, getting Mythos 5 back into roughly 100 enterprises and agencies limits the revenue damage from a near-total shutdown, but the precedent is double-edged: a model's reach is now a function of regulatory clearance, and a future dispute could flip access off again overnight. For the sector, the read is that frontier AI is being treated like a controlled technology, which advantages labs with the smoothest government relationships and the most defensible safety stories, and disadvantages anyone whose models can be cast as a security risk. The clean tell to watch is how fast Fable 5 comes back, because the speed of that decision will signal whether this regime is a light-touch, negotiable one or a hard gate that can throttle the cadence of capability releases the whole industry's growth depends on.

The Magnificent Seven Close Out Their Worst Month in Years as AI Doubt Sets In

• The Magnificent Seven lost roughly $2 trillion in market value in June, accounting for more than two-thirds of the S&P 500's decline for the month

• The group fell a median of about 9.7% in June, while the rest of the S&P 500 posted a median gain of about 0.3%

• Microsoft and Amazon each shed more than $350 billion, Apple and Alphabet roughly $300 billion each, and Nvidia and Tesla about $260 billion and $200 billion; Microsoft is on pace for its worst month since December 2000

• The Roundhill Magnificent Seven ETF is heading for its steepest monthly drop since launching in 2023, with investors pulling about $1 billion from the fund

• Why it matters: The mega-caps that carried the market for two years just had the breadth of the rally turn against them, a sign that AI enthusiasm is finally meeting price discipline.

June was the month the AI trade cracked. The Magnificent Seven, the seven mega-caps that drove the market to records, collectively lost about $2 trillion in market value over the month, more than two-thirds of the entire S&P 500's decline, according to data compiled by Yahoo Finance. The damage was concentrated and severe: the group fell a median of roughly 9.7% while the other 493 companies in the index posted a median gain of about 0.3%, a rare reversal of the narrow leadership that has defined the index for two years. (Yahoo Finance, Benzinga)

The individual scorecard underscores how broad the retreat was. Microsoft and Amazon each lost more than $350 billion in value, Apple and Alphabet roughly $300 billion apiece, Nvidia about $260 billion, and Tesla around $200 billion. Microsoft, the group's worst performer, is on track for its steepest monthly decline since December 2000. The Roundhill Magnificent Seven ETF, which bundles the seven into a single ticker, is set for its worst month since it launched in 2023 and saw about $1 billion in outflows. Underneath the numbers sits a single question that has hung over the tape all month: when do the more than $700 billion in 2026 AI infrastructure commitments from Microsoft, Amazon, Alphabet, and Meta actually start paying off. (Yahoo Finance, Benzinga)

Why it matters: A $2 trillion drawdown concentrated in seven names is the market's way of repricing the AI thesis from "spend whatever it takes" to "show me the return." The selloff is not about broken businesses; revenues and earnings at most of these companies are still growing. It is about the math of the buildout, with hyperscaler capex climbing toward 90% of operating cash flow this year while the revenue that capex is supposed to generate remains mostly a promise. That is why the day's other stories matter so much in this context: Samsung and SK Hynix committing $1.3 trillion to supply, Google unable to sell Meta enough compute, and the government gating who can use the best models all feed the same investor anxiety about how much capital this cycle consumes and how long until it returns cash. For portfolios, the breadth signal is the important one. When the median S&P name is up and the mega-caps are down nearly 10%, the market is rotating, not collapsing, and the question for the second half is whether that rotation is a healthy broadening or the early innings of the AI trade deflating. Either way, the era of buying these seven names without asking about the payoff appears to be over.