Matt Marshall✓

The AI bubble isn’t where Wall Street is looking. It’s inside the enterprise. 79% of companies running AI say the GPUs they use sit idle at least half the time. Yet 70% of those companies plan new AI infrastructure, according to three months of VentureBeat Intelligence surveys.

Even among companies running their own stack, 56% say the GPUs they use sit idle at least half the time. One reason: When an agent waits on a tool call or a user, GPU compute drops to zero while the session’s memory cache stays locked. Agents run out of memory before compute.

Not everyone sees a glut. @weka's @valb000 says demand for tokens now outstrips supply at any price, as security agents run around the clock. We’ll debate it Wednesday at our Agentic Infrastructure Stack Summit in SF. Apply by Monday:

America answered China’s lead in open AI models twice in the last 24 hours, with opposite bets. Meta’s Muse Spark Open is built for power. Reflection’s Beam, out today, was trained to use fewer tokens. On one coding test, it used about a third of the tokens of China’s GLM-5.2.

Reflection’s bet: Buyers will care more about the cost of each finished task than the price of each token. Nobody has published cost per task for these models yet, so we’re measuring it. My interview with Reflection’s technology chief:

Reflection’s Ioannis Antonoglou also made the case against running Chinese models on your own servers. You can’t be sure a model wasn’t “trained to just activate certain behaviors in certain scenarios,” he told me. Then: “This is not to say that Chinese models are bad.”

Companies pay for Claude. They run their agents on OpenAI. Ramp reports more businesses now pay Anthropic than OpenAI. Yet in VentureBeat’s September survey, twice as many companies run AI agents on OpenAI’s platform as on Anthropic’s: 43% to 22%.

Anthropic’s opening: Among 61 companies shopping for a new agent platform, 33% are considering Anthropic’s and 30% OpenAI’s, a statistically insignificant gap. More on the full @VentureBeat survey, including Google’s numbers:

Model switching is costly: 59% of companies that swapped the model in a live app had to rework the app or revert. Home Depot CIO Angie Brown: “One of the things is the context training that has to be done anytime there’s a switch, so there’s a stickiness level..."

Enterprises run several AI agent platforms at once. Only one gets named the primary. Of the 75 companies running OpenAI's agent platform, 52 named it the primary. That's 69%. Of the 45 running Anthropic's, 17 did. That's 38%. New VentureBeat Research. 169 enterprises.

Primary platform, full August ranking: OpenAI 33% Google 24% Microsoft 13% Anthropic 11% Enterprise apps (Salesforce SAP, Workforce, etc) 11% Amazon 3% Anthropic is fourth. Behind Google and Microsoft. (@VentureBeat data)

Caveats, because they matter. 169 respondents, all at companies with 100+ employees. Self-selected panel, not a probability sample. This wave skewed mid-market. 16% were at companies over 10,000 employees, against 51% the month before. August snapshot, not a trend.

Most enterprises think they’re scaling AI. In reality, they’re just creating a mess. New @VentureBeat research exposes the "Governance Mirage": 72% of orgs now juggle 2+ "primary" platforms, leaving them with zero centralized control. The "Security Irony" is real. 🧵 (1/4)

For those who want the full data dive immediately, here is the link:

The irony? 40% of orgs rely on model providers to secure the agents they provide. But with vendor opacity as the #1 hurdle, per our research, you can't audit what you can't see. That's why 34.3% of leaders are moving toward a hybrid control plane to reclaim sovereignty. (3/4)

The era of "blind trust" in AI vendors is ending. Organizations need independent security instrumentation, not a vendor-locked "mirage." Read the full @VentureBeat flagship report on how the largest employers are fighting to reclaim the AI Control Plane: