Why AI infrastructure is the new battleground

Welcome back to the OptimusEdge AI. Four companies are on track to spend more on AI infrastructure this year than the entire Apollo space program cost, adjusted for inflation. Whether that's foresight or overreach is genuinely up for debate and both sides have real numbers behind them.
___________________________________________________________
The Edge Upload: Today’s Insights

A car company is building the world's largest chip factory: Tesla, SpaceX, and xAI broke ground on Terafab, a planned $16.8 billion (initial phase) semiconductor complex outside Houston aiming for 1 terawatt of annual chip output Intel has joined as a manufacturing partner, and Musk has said combined demand from his companies already exceeds global chip supply.

Agentic AI spend is about to overtake chatbots: Gartner projects agentic AI spending will hit $201.9 billion in 2026 and overtake chatbot spending by 2027 a sign the compute demand curve is shifting from "answer a question" to "complete a task," which uses meaningfully more compute per interaction.

Physical AI is still small, but growing fast: the physical AI market — robots and machines that perceive and act in the real world is estimated at $3.5 billion in 2025, projected to reach $58.1 billion by 2035.

A chatbot answers a question and stops.

An agent keeps going checking its own work, calling tools, retrying failed steps, coordinating with other agents and every one of those steps is a separate inference call.

That compounding is showing up in the numbers: inference already consumed roughly half of all AI compute in 2025, and that share is projected to climb to two-thirds in 2026 and 75% by 2030.

Physical AI adds a second, different kind of demand entirely robots need models that process real-time sensor data (vision, touch, spatial awareness) fast enough to act on it safely, which is a much heavier compute load per decision than generating text.

This is a genuinely contested question, and reasonable people land on both sides.

The bubble case: the four largest hyperscalers are on pace to spend around $725–765 billion on AI infrastructure in 2026 roughly 2.4% of U.S.

GDP, above levels seen even in the dot-com cycle while AI company revenue remains a small fraction of that spend.

Critics also point to real duplication: multiple companies training similar models and bidding up the same constrained resources (GPUs, power, HBM memory), which looks less like eight companies solving eight different problems and more like eight companies solving the same one, expensively.

The real-demand case: data center vacancy remains near zero in the most active U.S.

markets, and JLL research indicates available capacity will stay scarce at least through 2027 the industry currently looks more constrained by its ability to build than by a lack of demand.

National programs like the CHIPS and Science Act and equivalents in Europe and Asia also create a demand floor that doesn't depend on any single company's revenue: governments are treating AI infrastructure as strategic capacity regardless of near-term ROI.

Both things can be true at once genuine long-term demand, and some amount of overbuilding in the short term. The honest answer right now is that nobody, including the people spending the money, actually knows the ratio yet.

NVIDIA remains the baseline but it's no longer the only story.

Every hyperscaler is now building its own custom silicon (Google's TPU, Amazon's Trainium, Meta's MTIA) specifically to reduce dependence on any single vendor.

Musk's Terafab is the most unusual entrant: a semiconductor fab built by a car company, a rocket company, and an AI company jointly, explicitly because Musk says existing suppliers "cannot possibly produce enough hardware" for Tesla's robots, SpaceX's satellites, and xAI's models combined.

And increasingly, governments themselves are becoming builders sovereign AI investment is projected to exceed $100 billion in 2026, a figure that was close to zero a few years ago (a thread we covered in more depth in our on-prem vs. cloud vs. hybrid issue).

This build-out needs people who can run it, not just build it.

NVIDIA's own certification program has expanded significantly for 2026, with tracks specifically for AI infrastructure and operations (NCP-AIO), physical AI, and agentic AI aimed at cloud, data center, and network administrators who need to deploy and manage these systems, not just data scientists training models.

On the ground, that's translating into real, current job openings: AI infrastructure engineer roles in the U.S. are averaging around $154,000/year, with a typical range of $113,000–$197,000 depending on experience and location.

For anyone already in networking or cloud infrastructure, this is less a career change than a natural extension CUDA, GPU orchestration, and distributed systems knowledge on top of skills most infrastructure engineers already have.

Takeaway: The compute demand behind agentic and physical AI is real and measurable inference alone is set to triple its share of AI compute by 2030. Whether current spending matches that demand exactly is genuinely unresolved. Either way, the infrastructure and the people who can run it are being built right now, not in some hypothetical future.

That's today's briefing. If your team's been asking "is this all just hype," forward them the bubble section it's the more honest answer than a yes or no. Past issues are in the archive. See you tomorrow

NVIDIA Deep Learning Institute: NVIDIA's own training platform, with self-paced and instructor-led courses covering AI infrastructure, DGX system management, and networking the practical starting point if the "new jobs" section above is relevant to you.

Terafab Details: Tesla and SpaceX will invest $16.8B to start building 'Terafab' chip factory in Texas TechCrunch's coverage of the formal announcement.

Compute Demand Data: Roundup of Agentic AI Forecasts and Market Estimates, 2026 the inference-share-of-compute figures, sourced from McKinsey

The Bubble Debate: Is the AI Infrastructure Build-Out a Bubble? Here's What the Data Actually Shows the GDP-share and valuation comparison to the dot-com era.

NVIDIA Certification Path: NVIDIA Certifications for AI, ML & Generative AI Careers in 2026 a breakdown of the current certification tracks, including AI infrastructure and operations

That’s it for today !☀

Edge AI is levelling up are you? Until next time, stay curious, stay building, and don’t let your machines take over. 🤖😆

Enjoyed today’s issue? Share OptimusEdge AI with your engineering team. Subscribe to OptimusEdge AI

Your Edge AI Explorer,
Sharat Sami (Let’s connect on LinkedIn)