Skip to main content
biznesam.ai
AI Briefing

AI infrastructure hits 'physical world' — $156B projects blocked

Data center moratoriums, agent pivot, and $10B China investment — signals that AI market is reorienting from 'bigger models' to 'better systems and constrained infrastructure'.

LV EN

TLDR

AI infrastructure has collided with the physical world — 48 data center projects blocked ($156B investment at risk), 27 US states implementing moratoriums. Meanwhile: OpenAI/Anthropic/Google pivoting from models to agents, DeepSeek doubling down on open-source with $10B. Signal for CEOs: AI market is reorienting from “bigger models” to “better systems + constrained infrastructure”. Your cloud roadmap assumptions about compute availability may be outdated.

Story 1: Data center “veto power” — regulation halts $156B investments

What changed: 48 data center projects were blocked or delayed in 2025, affecting $156 billion in potential investment. 27 US states are advancing data center legislation (moratoriums, energy cost regulation). Bernie Sanders and Alexandria Ocasio-Cortez introduced the Artificial Intelligence Data Center Moratorium Act — a nationwide halt on any project ≥20 megawatts until “strong national safeguards” are in place. Ben Thompson (Stratechery) calls this the “data center veto” — unlike globalization, AI infrastructure exists in the physical world and requires permission. This gives ordinary people veto power they didn’t have over offshoring.

Why it matters: Your AI project delivery risks are no longer just budget and technology — regulation and community opposition become the bottleneck. Cloud provider compute growth assumptions (that justified your 2025 roadmap) may be outdated. If your strategy assumes AWS/Azure/GCP can infinitely scale capacity in 2026-2027, you have risk exposure. Supply constraint means pricing power shifts to vendors.

What to do this month: Ask your IT/Cloud lead 2 questions: (1) What are our AI workload SLA commitments with cloud providers for the next 18 months? Do they guarantee capacity or just “best effort”? (2) Do we have a Plan B if our primary region faces delays? Multi-cloud is no longer just cost arbitrage — it’s business continuity. If you have hardware-intensive workflows (rendering, ML training), evaluate hybrid/on-prem backup options.

I expect: Cloud pricing for GPU/TPU instances will rise 15-25% over the next 12 months (supply constrained). Providers will start prioritizing enterprise contracts over spot/on-demand. Companies with fixed-price SaaS products (paying cloud per usage) will face margin squeeze. On-prem AI appliances (e.g., NVIDIA DGX stations) will revive as a risk hedge — not economically optimal, but delivery-guaranteed.

Story 2: Model labs become agent labs — product definition changes

What changed: OpenAI, Anthropic, Google — all three major AI labs are simultaneously pivoting from models to agents. OpenAI is advancing Codex agent with appshots and remote computer use. Anthropic expanded Claude auto mode to all Pro subscribers. Google released Gemini Spark as a “24/7 personal AI agent” for recurring tasks. Even AI21 Labs (historically a pure model shop) is shutting down model teams to focus on agent products. Latent Space podcast summarizes: “The standalone model is no longer the product. The product is model + harness + workflow + UI + memory + economics.”

Why it matters: Your AI procurement criteria are outdated if you’re still evaluating vendors by benchmark scores (MMLU, HumanEval). The next AI adoption wave won’t be “better chatbot” — it will be “done for you”. This changes what you’re buying: not API access to a model, but an integrated agent system. Vendors with the best model but no agent harness/memory/workflow orchestration will lose enterprise deals. Your vendor selection process must change too — it’s no longer enough to evaluate “how well it answers prompts”, you need to test “how well it executes a 10-step workflow without human intervention”.

What to do this month: Identify 1-2 workflows in your organization that repeat at least 5x weekly and contain 3+ steps (e.g., “receive inquiry → check CRM → draft response → send follow-up”). Evaluate whether any of your current AI tools (ChatGPT Enterprise, Claude, Copilot) can automate it as an agent, not just “assist at each step”. If not — you’re using a 2024 product in 2026. Ask your AI vendor: “Can your product trigger actions in my systems (CRM, email, Slack) without me clicking ‘confirm’ at every step?” If the answer is “not yet”, they’re behind the curve.

I expect: Q3-Q4 2026 will see “agent wars” between OpenAI Codex, Anthropic Claude agents, and Google Gemini Spark. The winner won’t be whoever has the best model, but whoever has the best integration ecosystem (API connectors to Salesforce, HubSpot, Slack, Gmail, etc.). Pricing models will shift from “tokens” to “completed tasks” or “workflows executed”. CEOs still paying per-token at the end of 2026 will be paying 2-3x more than competitors with task-based pricing.

Story 3: DeepSeek $10B funding — open-source gets stronger

What changed: Chinese AI lab DeepSeek is raising $10 billion in its first external financing round. Founder Liang Wenfeng promised investors: priority is open-source model development, not short-term commercialization. Goal — AGI (artificial general intelligence). Investors include the National AI Industry Investment Fund (state fund), Tencent, and IDG Capital. DeepSeek already released V4-Pro (1.6 trillion parameters) and V4-Flash (284B parameters) — both open-source with permissive licensing. Latent Space notes: DeepSeek’s permanent 75% discount makes frontier-class inference “too cheap to meter”.

Why it matters: Open-source models just received a $10B war chest. This means the OpenAI/Anthropic proprietary moat is eroding faster than planned. If you signed a 3-year OpenAI Enterprise contract in 2025 with premium pricing based on “they have the best model”, you have vendor lock-in risk. DeepSeek’s strategy (open-source + China backing) makes it possible that by 2027, open-source models are 90% of GPT-4.7/Claude Opus quality but cost 10% of the price. This puts pricing pressure on OpenAI/Anthropic — their differentiation will no longer be “model quality”, but “integration + support + compliance”. Enterprise buyers with strategic foresight can now negotiate better terms or hedge bets with a multi-vendor strategy.

What to do this month: Pull out your current AI vendor contracts and identify: (1) Are you locked into a single vendor API? (2) What are the exit costs if you want to switch in 12 months? (3) Is your code/prompts vendor-agnostic or tightly coupled to OpenAI/Anthropic-specific features? If the answer is “locked-in, high exit cost, vendor-specific” — you have strategic vulnerability. Build a proof-of-concept with an open-source alternative (DeepSeek V4, Llama 4, or self-hosted variant) for at least one use case. The goal isn’t “switch now”, but “know you can switch in 2027 if pricing becomes unfavorable”.

I expect: OpenAI/Anthropic will announce “enterprise tier” changes in Q4 2026 — likely base tier price cuts, but premium charges for “agent orchestration” and “guaranteed uptime”. Their margin defense will shift from “we’re the only ones with GPT-4 level model” to “we’re the only ones with 99.99% SLA + SOC2 + GDPR compliance + white-glove onboarding”. Buyers who only look at $/token will use open-source. Buyers who value total cost of ownership (compliance, support, integration) will stay with proprietary — but leverage in negotiations will be much higher.

Pattern today

Today’s three stories point to one major shift: AI market is reorienting from “bigger models” to “better systems + constrained infrastructure”. Data center moratoriums give cloud providers pricing power (supply constraint). Open-source $10B backing erodes proprietary model moat. Agent pivot changes product definition from “API call” to “completed workflow”.

CEOs still thinking in 2024 terms (“when GPT-5?”, “how much per token?”) are already behind. 2026 questions are: “Can my cloud vendor deliver the capacity they promised?”, “Can my AI tool execute tasks end-to-end or just assist?”, “Am I locked into a vendor that might be overpriced in 12 months?”.

Additional signal: NVIDIA no longer reports gaming revenue separately — now it’s all “Edge Computing” (PC, consoles, robotics, automotive). Clear signal: data center is the only priority, gaming GPU availability will decline.

Sources