The New Cathedrals (And the Lobster Inside)

From $600 billion datacenters to the $0.06 question you just asked your AI agent — the real economics of intelligence

By the OpenClaw Team · February 2026 · 15 min read

“When they said $50 billion for a plant, I said, ‘What the hell kind of plant is that?’” — Donald Trump, on Meta’s Hyperion datacenter, August 2025 [1]

TL;DR

In 2026, the five biggest tech companies will spend over $600 billion building AI infrastructure — more than 4× the entire US energy sector’s capital expenditure. Meta alone is building a datacenter in Louisiana that will cover more land than two Manhattan Central Parks. This article follows the money from these billion-dollar cathedrals of compute all the way down to the fraction of a cent it costs when your AI agent answers a question. If you’ve ever wondered why AI companies are spending like drunken pharaohs while the actual product feels almost free — this is where you’ll find the answer. And if you want to understand what an AI agent like OpenClaw actually costs to run each month, skip to the cost breakdown section. Spoiler: it’s less than your Netflix subscription.

Part 1: The Cathedrals

$600 Billion Buys a Lot of Concrete

Our team has been in infrastructure for years. Kubernetes clusters, cloud environments, serverless functions — the kind of plumbing that keeps modern software running. We thought we understood scale.

Then we read the 2026 earnings reports.

Amazon: $200 billion in capital expenditure this year [2]. Two hundred billion. That’s more than the GDP of Greece.

Google (Alphabet): $175–185 billion [3]. Up from $91 billion last year — nearly doubling in 12 months.

Meta: $115–135 billion [4]. The company that makes its money from ads on Facebook and Instagram is spending more on construction than most countries spend on their entire military.

Microsoft: approximately $100 billion [5]. Fueling Azure and its partnership with OpenAI.

Oracle: roughly $40 billion [5]. The database company quietly becoming a datacenter empire.

Combined: over $600 billion in a single year. That’s a 36% increase over 2025. Roughly 75% of it — about $450 billion — is going directly to AI infrastructure [6].

To put this in perspective: the combined 2026 capex of these five companies exceeds four times what the entire publicly traded US energy sector spends to drill exploration wells, extract oil and gas, refine fuel, and run chemical plants [7]. Amazon alone outspends the entire US energy sector.

We are witnessing the largest concentrated infrastructure buildout in human history. And most people have never seen any of it.

Hyperscaler Capital Expenditure 2026

Hyperion: The Datacenter That Ate Louisiana

The most dramatic example is Meta’s Hyperion datacenter in Holly Ridge, Richland Parish, Louisiana.

The original site was 2,250 acres — already enormous. Then, as Fortune reported in early February 2026, Meta quietly purchased another 1,400 adjacent acres, bringing the total campus to 3,650 acres [1]. That’s more than twice the size of New Orleans’ largest airport. When President Trump held up a graphic comparing Hyperion to Manhattan, it covered a significant portion of the island.

The project’s estimated cost is $27 billion, financed through a joint venture with Blue Owl Capital in what was described as the largest private-credit transaction ever [8]. Construction is already underway with 3,700+ workers on site, scaling to 5,000 [8]. The facility will eventually support up to 5 gigawatts of capacity — enough electricity to power over 3 million homes.

Why is a social media company building something that looks like it belongs in a science fiction movie? Because Meta isn’t just a social media company anymore. As their spokesperson stated, “This expansion is critical for our superintelligence models” [8].

Meanwhile, Meta is also building a separate 1-gigawatt datacenter in Indiana (cost: over $10 billion) and has signed cloud deals worth over $40 billion with Google, CoreWeave, and Oracle [3]. Zuckerberg described the company’s mission as delivering “personal superintelligence to billions of people” [5].

GPUs in Tents, Reactors in Fields

The urgency is creating some remarkable scenes.

When demand outpaced Meta’s ability to build permanent structures fast enough, the company deployed GPUs in temporary tent-like facilities — a project internally called Prometheus [9]. Rows of high-end processors, each costing tens of thousands of dollars, running inside what looked like glorified warehouses, because permanent buildings couldn’t be erected quickly enough.

Cooling is one of the biggest engineering challenges. xAI’s datacenter in Memphis runs traditional air and liquid cooling. Nebius (the AI cloud company spun out of Yandex) built its primary datacenter in Finland, using the country’s naturally cold climate as a free cooling system [9]. Meta’s Louisiana site faces the opposite problem — humid subtropical heat — requiring massive cooling infrastructure.

The most ambitious energy play: nuclear. Meta has deals with TerraPower and Oklo for nuclear power. Google signed with Westinghouse. Microsoft invested in nuclear through its Stargate partnership with OpenAI. The reasoning is straightforward — AI datacenters need baseload power that doesn’t depend on weather, and renewables alone can’t guarantee the 24/7 reliability these facilities demand [9].

OpenAI’s Stargate project, backed by a consortium including SoftBank, is planning for 1.2 gigawatts of demand. Satellite imagery shows construction sites with 6,000+ workers visible from space [9]. The scale is so large that local power grids are being redesigned to accommodate them.

Part 2: Training vs. Inference — Two Very Different Economies

All of this infrastructure serves two fundamentally different purposes, and understanding the distinction is key to understanding AI economics.

Training: The One-Time Investment

Training is the process of creating an AI model — feeding it vast amounts of text, code, and data, and adjusting billions of parameters until it can generate coherent, useful responses. Training a frontier model (GPT-5, Claude Opus, Gemini Ultra) costs $100 million or more and takes weeks or months on thousands of GPUs running simultaneously [10].

Think of training as building a factory. It’s enormously expensive, but you only do it once (per model version). Once the model is trained, it exists as a set of numerical weights — essentially a very large file — that can be copied and deployed anywhere.

Training is why the datacenters are so big. You need thousands of GPUs working in concert, connected by ultra-fast networking, consuming megawatts of power, for weeks at a time. The $600 billion in infrastructure spending is largely about training the next generation of models.

Inference: The Running Cost

Inference is what happens when you actually use the model — when you type a question and it generates an answer. Every time you send a message to ChatGPT, Claude, or an OpenClaw agent, you’re performing inference.

Here’s what’s remarkable: inference is absurdly cheap per query. We’re talking fractions of a cent for a typical question. But there are billions of queries happening every day across all AI services, and those fractions add up.

Here’s a useful analogy: training is like building a power plant. Inference is the electricity bill. The plant costs billions to construct, but the electricity flowing through it costs pennies per kilowatt-hour. Similarly, training costs hundreds of millions, but each individual inference costs a tiny fraction of a cent.

This separation is why AI companies can spend $600 billion on infrastructure while charging users $20/month for ChatGPT Plus. The infrastructure investment creates the capability. The per-query cost is what users actually pay.

Part 3: The Token Economy

To understand what AI agents actually cost to run, you need to understand tokens. This section is the one we wish someone had written for us when we started building GoClaw.

What’s a Token?

A token is the fundamental unit that AI models process. It’s not exactly a word — it’s a chunk of text, typically 3–4 characters or about ¾ of a word in English [11]. “Hello, how are you?” is about 6 tokens. A typical email might be 200–400 tokens. A full page of text is roughly 500–700 tokens.

AI providers charge per token, separately for input (what you send) and output (what the model generates). This distinction matters enormously.

Why Output Costs 3–5× More

Here’s something that surprises most people: output tokens are 3 to 5 times more expensive than input tokens [12].

The reason is computational. Input tokens are processed in parallel — the model reads your entire prompt at once using matrix operations that GPUs are very good at. Output tokens are generated sequentially — one at a time, each depending on all the tokens that came before it. Sequential generation is much more computationally expensive than parallel processing.

This is why pricing pages can be misleading. When you see “GPT-4o: $2.50 per million tokens,” that’s the input price. Output is $10.00 per million — four times higher [13]. For a typical chatbot interaction where the model generates 2× more tokens than you send, your actual cost is dominated by the output side.

The Price Landscape (Early 2026)

The market has stratified into clear tiers:

Budget tier — DeepSeek V3 ($0.28/$1.10 per million tokens), Gemini Flash Lite ($0.08/$0.30), GPT-4o Mini ($0.15/$0.60). These models handle 70–80% of routine tasks at near-zero cost [14].

Mid tier — Claude Sonnet 4 ($3/$15), GPT-4o ($2.50/$10), Gemini 2.5 Pro ($1.25/$10). The workhorse models that balance quality and cost [13].

Premium tier — Claude Opus 4.6 ($15/$75), GPT-5.2 Pro ($21/$168). Maximum intelligence for tasks where quality matters more than cost [13].

The spread is staggering. The cheapest model (Gemini Flash Lite) costs roughly 1,000 times less than the most expensive (GPT-5.2 Pro) per token. And here’s the kicker: for many everyday tasks, the cheap models perform indistinguishably from the expensive ones [14].

Token Pricing Across Model Tiers

The Speed of Decline

Perhaps the most important economic trend in AI: inference costs are falling 10× per year [15]. Not 10%. Ten times. Every year. GPT-4 equivalent performance that cost $20 per million tokens in late 2022 now costs about $0.40 [15]. That’s a 50× reduction in three years.

DeepSeek’s entry into the market in early 2025 was the earthquake. By offering performance comparable to GPT-4 at 90% lower prices, it triggered a price war that forced every major provider to slash costs [16]. Anthropic responded by dropping Claude Opus 4.5’s pricing by 66% compared to Opus 4 [17]. Google’s free tier for Gemini makes prototyping essentially cost-free.

The trajectory is clear: the cost of intelligence is plummeting. What costs a dollar today will cost a dime next year and a penny the year after. This is the most important chart in AI economics, and most people haven’t internalized it yet.

Part 4: What an AI Agent Actually Costs

Now let’s connect the cathedrals to the lobster. What does it actually cost to run an OpenClaw agent month-to-month?

An AI agent is more expensive per interaction than a simple chatbot, for a fundamental reason: an agent makes multiple AI calls per task. A chatbot receives your message and sends one response. An agent might:

1. Load context (your personality file, preferences, daily notes, skill list)
2. Interpret your request and select the right skill
3. Execute tool calls (web search, email check, calendar lookup)
4. Synthesize the results
5. Compose a reply
6. Save relevant information to memory

Each of these steps consumes tokens. A single complex task might involve 13,000+ input tokens and 1,000+ output tokens across multiple model calls [18].

Agent Cost Anatomy

The Math: What You’d Actually Pay

Let’s run the numbers for three usage profiles, using Claude Sonnet 4 ($3/$15 per million tokens) as the default model — the sweet spot of quality and cost that most OpenClaw users settle on:

Light user (20 tasks/day, simple queries): ~600K input + ~50K output tokens/day. Monthly cost: $3–5 in API fees.

Power user (50 tasks/day, mix of simple and complex): ~2M input + ~200K output tokens/day. Monthly cost: $10–25 in API fees.

Heavy user with multi-agent (100+ tasks/day, sub-agents, web browsing): ~5M input + ~500K output tokens/day, with some tasks routed to Opus. Monthly cost: $50–150 in API fees.

But API fees are only part of the picture. You also need infrastructure to run the Gateway:

Self-hosted VPS: $15–50/month for a decent virtual server (4GB+ RAM, persistent storage).

Self-hosted at home: A Mac Mini or old laptop with Node.js — essentially free if you already have one, but you’re paying in complexity, maintenance, and the anxiety of wondering what happens when your home internet goes down.

GoClaw: One predictable price that includes the infrastructure, the deployment, the security hardening, managed updates, and optionally the API access itself [19]. You don’t manage servers, you don’t juggle API keys, and you don’t debug Node.js dependency conflicts at 2 AM.

The Model Routing Trick

Here’s something that’s obvious once you see it but that most people miss: you don’t need to use the same model for every task.

A question like “What time is my meeting tomorrow?” doesn’t need Claude Opus at $75 per million output tokens. GPT-4o Mini at $0.60 handles it perfectly. But “Analyze this 50-page contract and identify the three most concerning clauses” absolutely benefits from Opus-level reasoning.

OpenClaw’s multi-agent architecture makes this natural. The orchestrator agent can route different sub-tasks to different models:

• Quick factual lookups → GPT-4o Mini or Gemini Flash ($0.15–0.60/M out)
• Routine tasks → Claude Sonnet or GPT-4o ($10–15/M out)
• Complex reasoning → Claude Opus ($75/M out)

One analysis found that routing 40% of queries to cheaper models while preserving quality reduced per-task costs from $0.15 to $0.054 — a 64% savings [14]. With smart routing, you get premium intelligence when it matters and economy pricing when it doesn’t. This is something GoClaw optimizes automatically.

Comparison: Agent vs. Human Assistant

Let’s put this in business terms:

Human VA

AI Agent (mid-tier)

AI Agent (premium)

Hourly cost

$15–30/hr

~$0.02/hr amortized

~$0.10/hr amortized

Availability

8 hrs/day, weekdays

24/7/365

24/7/365

Response time

Minutes to hours

Seconds

Seconds

Simultaneous tasks

1

Multiple (sub-agents)

Multiple

Learns your preferences

Slowly, inconsistently

Systematically, in writing

Systematically

Makes mistakes

Yes

Yes (different kinds)

Yes (fewer)

Monthly cost (moderate use)

$1,200–2,400

$15–40

$60–200

The AI agent isn’t better at everything — a human VA handles ambiguity, emotional intelligence, and novel situations far better. But for the 80% of tasks that are routine, structured, and information-heavy, the economics are overwhelming.

Part 5: Where the Money Goes (And Where It’s Going)

The Infrastructure Paradox

There’s a beautiful paradox at the heart of AI economics: companies are spending more money than ever on infrastructure, while the cost of using AI is falling faster than any technology in history.

How is this possible? Because demand is growing even faster than costs are falling. OpenAI reportedly serves over 400 million weekly active users [20]. Every query is cheap, but 400 million users asking multiple questions per day adds up to enormous compute demand. The $600 billion in infrastructure spending isn’t wasteful — it’s barely keeping up.

For the end user, this is wonderful news. The competition between providers, the open-source pressure from DeepSeek, the efficiency improvements in model architectures — all of this conspires to make intelligence cheaper and more accessible every quarter.

The Open Source Pressure Valve

DeepSeek deserves special mention because its impact on the market has been seismic. When a Chinese lab releases a model that matches GPT-4 performance at 90% lower cost, it puts a hard ceiling on what Western providers can charge [16]. It’s the equivalent of a generic drug entering a pharmaceutical market — suddenly, the premium brand has to justify its pricing with more than just “we were here first.”

This pressure is healthy. It means that even if you’re building on Claude or GPT today, the overall cost trend is permanently downward. Any business plan built on AI costs should assume costs will be 5–10× lower within two years. If that makes your business case marginal today, it makes it compelling tomorrow.

What This Means for GoClaw

This economic landscape is exactly why GoClaw exists and why we believe it will scale.

The raw ingredients — AI models, compute infrastructure, messaging APIs — are commoditizing. Every quarter, the same quality of AI gets cheaper. The same performance of cloud compute gets more affordable. But the assembly of these ingredients into a useful, secure, reliable personal agent remains hard.

GoClaw’s value proposition isn’t selling expensive AI. It’s packaging cheap AI into something that works. Every instance runs on Google Cloud with full isolation, encrypted secrets, and managed updates [19]. We can leverage model routing to optimize costs automatically — using premium models only when tasks demand it. As underlying costs continue to fall, the margin between what it costs us to serve an agent and what it costs a user to self-host (in time, expertise, and opportunity cost) only grows.

The $600 billion being spent by hyperscalers is, in a sense, GoClaw’s subsidy. They’re building the cathedrals. We’re helping people walk through the door.

The Takeaway

Here’s what we want you to remember:

The big numbers are real. $600 billion. Five gigawatt datacenters. Nuclear power plants. This isn’t hype — it’s the largest infrastructure buildout in history, happening right now, funded by the most profitable companies on earth.

The small numbers are also real. A fraction of a cent per question. $15–40 per month for a personal AI agent that works 24/7. Less than your coffee budget.

The gap between these numbers is the opportunity. Someone has to take the output of those hundred-billion-dollar datacenters and make it useful for the individual professional, the small business owner, the student, the founder. That’s what OpenClaw does at the software layer. That’s what GoClaw does at the infrastructure layer.

And the trajectory only goes in one direction. Costs are falling 10× per year. The agent that costs you $30/month today will cost $3/month in two years — and be smarter. Every month you wait is a month of compounding advantage you’re leaving on the table.

The cathedrals are built. The lobster is inside. Your turn. 🦞

👉 Start your AI agent today at goclaw.io. One price, everything included, 60 seconds to launch.

This is the third article in a series about the AI agent revolution. Previous: “The Lobster That Broke the Internet” and “Software That Rewrites Itself”. Next up: “Your AI Employee Starts Monday” — a practical guide to setting up your first agent, choosing the right model, and the first 10 tasks to delegate.

References

[1] Fortune, “Meta is quietly expanding its $10 billion Hyperion AI data center, now sprawling to four times the size of Manhattan’s Central Park,” Feb 4, 2026. Link

[2] Yahoo Finance, “Meta announces plans to build 1-gigawatt data center in Indiana,” Feb 11, 2026. Meta: $115-135B capex 2026; Amazon: $200B. Link

[3] Sherwood News, “How the Big Tech companies are spending their huge capex budgets,” Feb 10, 2026. Google: $175-185B. Link

[4] Meta Q4 2025 earnings: capex guidance $115-135B for 2026. Link

[5] Sherwood News, ibid. Microsoft ~$100B (analyst estimates), Oracle capex. Link

[6] IEEE ComSoc Technology Blog, “Hyperscaler capex > $600 bn in 2026, a 36% increase over 2025.” ~75% ($450B) for AI infrastructure. Link

[7] Morningstar, “AI Arms Race: How Tech’s Capital Surge Will Reshape the Investment Landscape in 2026.” Combined capex > 4× US energy sector. Link

[8] ROIC News, “EY Flags Meta’s $27 Billion Data-Center Accounting as High Risk,” Feb 11, 2026. 3,700+ workers, $27B JV with Blue Owl. Link

[9] Research compiled from multiple sources on datacenter construction trends, nuclear deals, and cooling strategies. Various industry reporting, Q4 2025–Q1 2026.

[10] Industry estimates for frontier model training costs. Various sources including Epoch AI, SemiAnalysis.

[11] OpenAI documentation: “A helpful rule of thumb is that one token generally corresponds to ~4 characters of text for common English text.” Link

[12] Silicon Data, “Understanding LLM Cost Per Token: A 2026 Practical Guide.” Output tokens 3-8× more expensive due to sequential generation. Link

[13] IntuitionLabs, “AI API Pricing Comparison (2026).” Comprehensive pricing across providers. Link

[14] CloudIDR, “Complete LLM Pricing Comparison 2026.” Model routing reducing costs 40-60%; 70-80% of workloads served by mid-tier models. Link

[15] Introl Blog, “Inference Unit Economics: The True Cost Per Million Tokens.” LLM inference costs declined 10× annually. Link

[16] IntuitionLabs, “LLM API Pricing Comparison (2025).” DeepSeek R1 debuted at 90% below Western competitors. Link

[17] Swfte AI, “AI API Pricing Trends 2026.” Claude Opus 4.5: 66% price reduction from Opus 4. Link

[18] Author’s estimates based on OpenClaw token consumption analysis and agent task profiling.

[19] GoClaw — hosted OpenClaw instances with managed infrastructure. Link

[20] OpenAI reported user figures, various sources, early 2026.

‍

‍