By The fact that the world's biggest energy conference is now headlined by cloud infrastructure executives tells you everything about where the money is going in 2026.
CERAWeek by S&P Global kicks off today in Houston, running March 23-27. For decades, this was the conference where oil executives and energy ministers set the agenda. This year, the speaker list reads like a hyperscaler org chart: Ruth Porat (Alphabet President and Chief Investment Officer), Brad Smith (Microsoft Vice Chair and President), Kerry Person (AWS VP of Data Center Planning and Delivery), Amanda Peterson Corio (Google's Global Head of Data Center Energy), Marc Spieler (NVIDIA's Senior Managing Director for Global Energy), and Patrick Ryan (Meta's Principal of Energy Strategy).
That is not a sideshow panel. Those are the headliners. At an energy conference.
The numbers driving the convergence
The reason these executives are in Houston, not at re:Invent or Google Cloud Next, is electricity. Training and running AI models requires staggering amounts of power, and the hyperscalers are collectively projected to spend over $600 billion in capital expenditures in 2026, according to Data Center Knowledge, citing Synergy Research Group data. Amazon alone has projected $200 billion in 2026 capex spending, up from $131 billion in 2025, per TechCrunch. Google is not far behind, with estimates between $175 billion and $185 billion, roughly double its 2025 spending.
These are not R&D budgets. This is concrete, copper, and silicon. According to Synergy Research Group, 1,297 hyperscale data centers are currently operational worldwide, with a pipeline of 770 more at various stages of planning and construction. Total hyperscale capacity is expected to double in just over 12 quarters.
CERAWeek 2026's theme is "Convergence and Competition: Energy, Technology and Geopolitics." Daniel Yergin, conference chair and Vice Chairman of S&P Global, said in the conference announcement: "The energy and technology industries are converging like never before."
What the chip deals tell us
Four days before CERAWeek opened, Nvidia and AWS confirmed a deal for 1 million GPU chips, with sales starting this year and extending through 2027. Ian Buck, Nvidia's VP of Hyperscale and High-Performance Computing, told Reuters the deal goes well beyond GPUs, including Nvidia's Spectrum networking chips and the Groq chips Nvidia released after its $17 billion licensing deal with an AI chip startup late last year.
The detail that matters for developers: AWS plans to combine Nvidia's Groq chips with six other Nvidia chip types for inference workloads. Buck told Reuters: "Inference is hard. It's wickedly hard. To be the best at inference, it is not a one chip pony. We actually use all seven chips."
That last line is the most important thing said about cloud infrastructure this month. If you're building on AWS and running inference-heavy workloads, the underlying hardware is about to get significantly more heterogeneous. The days of "pick a GPU instance and go" are ending. AWS is also adopting Nvidia's ConnectX and Spectrum X networking gear in its data centers, a notable shift for a company that has historically built its own custom networking stack.
Meanwhile, Microsoft is deploying custom Azure Maia 100 AI accelerators and Cobalt 100 CPUs, interconnected by 120,000 miles of dedicated fiber on its AI Wide Area Network, according to Data Center Knowledge. Google is expanding to 42 cloud regions with 127 Availability Zones. The chip strategy divergence across providers is real, and it has direct implications for which workloads run best where.
The vendor lock-in trap nobody is talking about
Here is the part that should concern any engineering team making stack decisions right now: as each cloud provider optimizes for different chip architectures and networking topologies, the portability story gets worse, not better.
AWS is going deep on Nvidia's full stack. Microsoft is betting on custom silicon (Maia, Cobalt) plus Nvidia. Google has its TPUs plus Nvidia partnerships. Each path means different instance types, different performance profiles for inference and training, and different optimization patterns in your code. The time-to-hello-world for an AI inference workload varies wildly depending on which provider's hardware assumptions your model was optimized for.
If you're a CTO evaluating cloud providers for AI workloads in 2026, the question is no longer "who has the most GPUs" but "whose chip strategy aligns with my inference patterns, and how painful will it be to leave?"
The power problem is the real bottleneck
The Houston Chronicle called 2026 the "prove it" year for AI infrastructure investment. That framing is right. The hyperscalers have committed hundreds of billions, but execution depends on something none of them fully control: power grid capacity.
Data Center Knowledge reported that as of mid-2025, more than 36 data center projects representing $162 billion in investment were either blocked or significantly delayed. Communities are pushing back. A Politico poll found that 25% of respondents could not correctly identify what a data center does, while others are frustrated by rising power costs regardless of whether data centers are the primary cause.
This is why the CERAWeek speaker list matters. AWS, Google, Microsoft, and Meta did not send their AI researchers to Houston. They sent their energy and data center infrastructure leads. Google alone has four speakers focused on energy: Amanda Peterson Corio (Data Center Energy), Ruth Porat (Chief Investment Officer), Raiford Smith (Power and Energy), and Michael Terrell (Advanced Energy). Microsoft sent five energy-focused speakers.
The bottleneck for AI infrastructure in 2026 is not chips. It is watts.
What this means for your stack decisions
If you are making infrastructure bets right now, here is what CERAWeek and the surrounding deals signal:
Inference is the new battleground. The Nvidia-AWS deal is explicitly oriented around multi-chip inference optimization, not training. Every major provider is shifting investment toward serving AI workloads, not just building them. Your inference cost per token will vary dramatically by provider this year.
Custom silicon fragmentation is accelerating. AWS (Trainium, Graviton, plus Nvidia's full stack), Azure (Maia, Cobalt, plus Nvidia), and GCP (TPUs plus Nvidia) are all diverging further. Multi-cloud AI is getting harder, not easier.
Power constraints will affect availability. If your preferred region is capacity-constrained because the local grid can't support new data center construction, you may face longer wait times for GPU instances or get pushed to less optimal regions. This is already happening.
The energy-tech convergence is permanent. CERAWeek creating "The Bridge," a new venue explicitly connecting energy and technology leaders, signals this is not a trend. The companies building your cloud infrastructure now think of themselves as energy companies first and compute companies second.
For developers, the practical takeaway: benchmark your inference workloads across providers now, before the chip architectures diverge further and migration costs compound. The window for low-cost portability between AI cloud platforms is closing.
Leon Vasquez covers developer tools and cloud infrastructure for The Daily Vibe.
