Google limits Meta's Gemini use after capacity shortages

• rohanpaul_ai's FT summary says Google capped Meta's Gemini usage after Meta asked for more capacity than Google could supply.
• The reported constraint was physical infrastructure, not contract wording: the FT summary ties the limit to scarce chips, power, and networking, while testingcatalog's recap says the cap delayed some Meta work.
• Meta was reportedly using Gemini across safety automation, customer support, ad tools, coding, and other internal workflows, according to rohanpaul_ai's post.
• TechCrunch's Micron report and Epoch AI's compute estimate add the wider backdrop: AI demand is now fighting over HBM, DRAM, and data center buildout as much as over model access.

You can read the Financial Times report, skim rohanpaul_ai's summary for the specific Meta use cases, and jump to Epoch AI's estimate for the weirder macro point that frontier labs still use only a minority of global AI compute. TechCrunch's Micron piece fills in the hardware side, where memory supply has become part of the same bottleneck.

The core report is simple: Google reportedly restricted Meta's Gemini usage after Meta requested more model capacity than Google could provide. testingcatalog's recap and the reply-thread duplicate both point back to the same FT claim.

What makes the item interesting is who is constraining whom. Meta is one of the few companies with enormous in-house AI infrastructure, but the FT summary still describes it as leaning on a rival's model capacity for internal work.

rohanpaul_ai's post says Google's own problem is allocation: cloud customers, first-party Gemini products, and internal demand were all competing for the same compute estate. The same post adds that Sundar Pichai had already said capacity shortages held back Google Cloud growth and pushed backlog sharply higher.

That fits the broader supply story in TechCrunch's Micron report, which frames HBM and adjacent memory as the scarce layer sitting beside every GPU. Epoch AI's estimate makes the bottleneck look even stranger: global AI compute is already widely distributed, but the most in-demand model capacity is still locally constrained.

The reported impact was not limited to one prototype team. According to rohanpaul_ai's summary of the FT report, Meta had been using Gemini in:

• safety automation
• customer support
• ad tools
• coding
• other internal workflows

testingcatalog's recap adds that the cap reportedly caused delays in customer support and content moderation work. That detail matters because it places Gemini inside operational systems, not just experimentation.

The visible evidence trail is narrow so far. rohanpaul_ai's post cites the FT report directly, while testingcatalog's post and its thread duplicate are downstream summaries rather than new reporting.

That leaves one concrete public fact pattern: a rival model provider reportedly imposed usage limits, and the named reason was infrastructure scarcity. The rest of the story, including how much capacity Meta lost and whether it shifted load elsewhere, is not surfaced in the evidence here or in the linked Financial Times report summary text carried into the tweet.