AI Companies Are Burning Cash. The Ones Building Roads for AI Are Printing It.

Same Week, Two Companies Started Building for Non-Human Customers

On February 11, 2026, Coinbase launched Agentic Wallets. A crypto wallet designed specifically for AI agents. Agents can autonomously spend, earn, and trade cryptocurrencies without human intervention.

The same day, Stripe released a preview of Machine Payments. A new payment method allowing AI agents to pay for digital services directly using USDC stablecoin.

Stripe's Head of Product, Jeff Weinstein, said: "Autonomous agents are a new and crucial customer category for us."

But this article isn't about how clever Coinbase and Stripe are. It's about a deeper question: in this new customer category, who actually makes the money?

Everyone's instinct says: AI companies make the most. OpenAI, Anthropic, Google DeepMind built the agents, so they should be the biggest winners.

The data says otherwise.

---

In a Gold Rush, the Biggest Winners Aren't the Miners

In the 1848 California Gold Rush, the people who got systematically rich weren't the miners panning for gold. It was Levi Strauss (selling jeans), Wells Fargo (providing transport and banking), and Samuel Brannan (buying up all the shovels and marking them up). Miners took the most risk and did the hardest labor, but systemic wealth flowed to infrastructure providers.

The 2025-2026 AI gold rush is replaying the same script.

Coinbase isn't an AI company. It's an exchange. But if an AI agent needs to spend money, it needs Coinbase's wallet. Stripe isn't an AI company. It's a payment processor. But if an AI agent needs to pay for something, it needs Stripe's rails. Cloudflare isn't an AI company. It's a CDN and edge computing platform. But if an AI agent needs to access the internet, it needs Cloudflare's pipes. Shopify isn't an AI company. It's an e-commerce platform. But if an AI agent needs to buy something, it needs Shopify's product catalog.

What these companies share: they don't make AI, but every economic activity of an AI agent must pass through them.

The financial data validates this thesis.

Cloudflare Q4 2025 revenue hit $614.5 million, up 34% year-over-year. It landed an $85 million contract from one AI company allocating 100% of its traffic to Cloudflare. Largest single ACV: $42.5 million. 80% of leading AI companies use Cloudflare infrastructure.

Stripe processed $1.4 trillion in 2024 payments, valued at $106.7 billion. OpenAI and Anthropic are both customers.

Shopify's AI-search orders grew 15x since January 2025.

Meanwhile, AI model companies are locked in a brutally margin-compressive competition. OpenAI projected $5 billion in losses for 2024. Anthropic burns cash even faster. DeepSeek replicated comparable capabilities at far lower cost, directly squeezing pricing power for all closed-source models. AI models are rapidly commoditizing.

Infrastructure doesn't commoditize. Whether GPT, Claude, Gemini, or open-source models win, AI agents all need payments, need networks, need product catalogs. Shovel sellers don't care who finds the gold.

---

HTTP 402 Waited 27 Years. Not for Humans.

In 1997, the architects of the HTTP protocol reserved a status code: 402 Payment Required. They envisioned a future where the web would have a native payment layer.

That status code sat dormant for 27 years. Nobody used it. Human payment flows require interfaces, confirmation buttons, credit card forms, redirects to third-party checkout pages. "Native HTTP payments" made no sense for humans.

Then AI agents appeared.

AI agents don't need interfaces. They don't fill forms. They don't click buttons. They need: send an HTTP request, receive a 402 response with payment details, pay with stablecoins, resend the request for data. Milliseconds.

In May 2025, Coinbase and Cloudflare jointly launched the x402 protocol, officially activating the dormant status code. Within months: 15 million transactions. By the December 2025 V2: over 100 million. By February 2026: over 50 million cumulative. The x402 Foundation was established. Google integrated x402 into its Agentic Payments Protocol. Stripe built Machine Payments on x402.

27 years ago, humans designed a payment protocol for machines. 27 years later, the machines showed up.

This is the shovel thesis in action. x402 isn't an AI model. It's the AI agent's payment pipe. No matter which AI wins, agents need to pay for APIs. x402 is the toll booth.

---

Salesforce Fired 4,000 People, Saved $100 Million, Then Sold That Same AI to 18,500 Customers

This is the case study worth thinking about the hardest.

In 2025, Salesforce deployed its own AI agent platform, Agentforce, internally. The result: AI agents took over 50% of customer service conversations. The support team was cut from 9,000 to 5,000 people. Annual savings: $100 million. CEO Marc Benioff personally announced this at Dreamforce 2025.

Then Salesforce did something even more interesting: it packaged the same technology as a product and sold it to other companies.

Agentforce closed 18,500 deals in its first year, 9,500 of them paid. 8,000+ customers running in production. By early 2026 (FY26 Q3), Agentforce ARR reached $540 million, up 330% year-over-year. The broader AI and Data business hit nearly $1.4 billion ARR, up 114%. 3.2 trillion tokens processed through the LLM gateway.

Pricing model: from traditional per-seat to Flex Credits. Each AI agent action consumes 20 credits, approximately $0.10. 100,000 credits cost $500.

Here's an original calculation.

A mid-size enterprise has 100 AI agents running (replacing 100 human support seats). Under traditional per-seat pricing, Enterprise Edition at ~$300/seat/month: 100 seats = $30,000/month = $360,000/year.

Under per-action pricing: each agent handles 200 conversations per day, with an average of 5 actions per conversation. 100 agents × 200 conversations × 5 actions = 100,000 actions/day. At $0.10/action = $10,000/day = $3.65 million/year.

Same number of "users." Revenue: 10x higher.

That's why Salesforce is willing to dismantle the per-seat model it invented. It's not cutting prices. It's switching to a billing model with a far higher ceiling. When your customer is a machine, the machine doesn't clock out at 5 PM. It runs 24/7, and every action generates revenue.

Salesforce ate its own dog food, proved the system works, then sold the recipe to 18,500 customers. This isn't a "To AI" slogan. It's a completed transaction.

---

Four Protocols Are Making AI Agents Into Economic Entities

If you think the above is just "one company's story," step back to the bigger picture. Four new protocols are building a complete economic infrastructure for AI agents.

MCP (Model Context Protocol) — How AI calls tools and data. Released by Anthropic, adopted by OpenAI and Google. AI's "USB-C port."

A2A (Agent2Agent Protocol) — How AI agents talk to each other. Released by Google Cloud, 50+ partners (Atlassian, PayPal, Salesforce, SAP). Open-sourced under the Linux Foundation.

x402 — How AI agents pay for things. Native HTTP payment layer. Driven by Coinbase and Cloudflare.

UCP (Universal Commerce Protocol) — How AI agents buy things. Co-released by Shopify and Google. End-to-end from product discovery to post-purchase.

Put them together: AI agents can use tools (MCP), communicate with peers (A2A), pay for services (x402), and purchase goods (UCP). That's every fundamental capability a human economic participant has.

Now look at who built these protocols. Coinbase (exchange), Cloudflare (CDN), Shopify (e-commerce), Google (search/ads), Stripe (payments). Except for Anthropic (MCP's originator), they're all infrastructure companies.

The protocol builders are the toll booth builders. Same as the TCP/IP era. Cisco didn't make content. But all content had to pass through Cisco's routers.

---

Who Gets Destroyed

Talking about winners isn't enough. Deep analysis must identify the losers.

Publishers: Search Traffic Down by a Third in One Year

Chartbeat data shows that through November 2025, global publisher Google search traffic declined by one-third in a single year.

This isn't a forecast. It already happened.

The cause: Google's AI Overviews. When users search a question, Google generates an AI answer directly at the top of results. Users get information without clicking any link. In 2026, over 80% of searches end with zero clicks.

Ahrefs' analysis is more specific: when AI Overviews appear, the #1-ranked page's CTR drops from 7.3% to 2.6%, a 34.5% decline. Paid ad CTR drops up to 53.6%.

Ad network Raptive estimated that full AI Overview rollout will reduce search traffic by 25% across its 5,000 publishers, a combined loss of approximately $2 billion in annual ad revenue.

News executives surveyed are even more pessimistic: they expect search traffic to decline by an average of 43% within three years, with some expecting losses exceeding 75%.

The SEO Industry: From Optimizing for Humans to Optimizing for Machines

Traditional SEO's core assumption: humans use keywords in search engines, and your job is to rank first. That assumption is crumbling. AI-driven referral traffic grew 527% year-over-year (Semrush data), albeit from a small base, growing far faster than traditional search.

A new discipline is replacing SEO: AEO (Answer Engine Optimization). SEO optimizes for "ranking." AEO optimizes for "being cited." The new KPI isn't "search ranking." It's "Share of Model" (how frequently AI recommends your brand).

New job titles are emerging: AI-SEO Strategist, Semantic Content Architect, AI Tool Integrator. The old roles (bulk content producers, backlink builders, basic keyword researchers) are being automated away by AI itself.

Digital Advertising: AI Agents Don't Watch Ads

This is the largest structural risk.

When the consumer journey shifts from "search on Google → see ad → click → visit website → buy" to "tell AI to buy shoes → AI agent auto-compares and orders," the entire logic chain of digital advertising breaks.

AI agents don't see banner ads. AI agents aren't moved by brand stories. AI agents decide based on structured data, price, ratings, and inventory.

By 2030, autonomous AI agents are projected to manage 80% of digital media buys. The ad budget you spend on Google Ads today will soon be spent by AI agents making media decisions for you. The target audience of advertising shifts from "human consumers" to "AI agents making decisions for humans."

Google itself sees this. Q4 2025 search ad revenue was $63.07 billion, still growing 17%. But Google Network revenue (third-party publisher sites) has started declining. Google's fix: embed ads into AI Overviews and AI Mode. Whether that works remains uncertain.

---

Your Store No Longer Needs a Beautiful Website

At NRF 2026, Shopify and Google co-released UCP (Universal Commerce Protocol). Consumers can shop directly within Google Gemini, Microsoft Copilot, and ChatGPT conversations. No browser. No website visit. No webpage needed.

Shopify calls these "Agentic Storefronts." Google calls its search-embedded virtual sales associate "Business Agent."

Shopify also launched an "Agentic Plan" open to all brands (not just Shopify merchants), letting any brand connect product catalogs to the Shopify Catalog for AI agent discovery and transactions.

A brand used to need a beautiful website, SEO strategy, ad spend, and conversion funnels. Now it needs: structured product data, API-accessible inventory, standardized checkout flows. Not a webpage for humans to view. Data for AI to read.

Through the gold rush lens: Shopify isn't an AI company, but it's becoming AI agents' "product catalog." No matter which AI agent runs on Gemini or ChatGPT, it queries Shopify's Catalog when buying things.

---

40% of Projects Will Fail, But Shovels Won't Go Unsold

Gartner's cold data: over 40% of agentic AI projects will be canceled by 2027. Of thousands of vendors claiming agentic capabilities, only about 130 have genuine ability. The rest is "Agent Washing."

But note a critical distinction: Gartner is talking about "agentic AI projects," not "companies providing infrastructure for AI agents."

Back to the gold rush analogy. In the 1848-1855 California Gold Rush, most miners came away with nothing. But Levi Strauss's jeans sales didn't decline because of miner failure rates. Whether miners found gold didn't affect their need to buy shovels and pants.

40% of agentic AI projects failing won't reduce AI agents' demand for payment rails (Stripe/Coinbase), network infrastructure (Cloudflare), product catalogs (Shopify), or enterprise runtime environments (Salesforce).

The directional data that matters:

AI agent market: $5.4 billion (2024) → $50.3 billion (2030), CAGR 45.8%.

By 2028: 15% of daily work decisions made autonomously by agentic AI (up from 0% in 2024). 33% of enterprise software will include agentic AI (up from less than 1% in 2024).

IDC's projection is more aggressive: over 1 billion AI agents globally by 2029. A 40x increase from 2025.

Every single agent needs network connectivity, payment capability, and data access. Shovel demand only grows.

---

Your Next Most Important Customer

Everyone is chasing AI model companies. NVIDIA, OpenAI, Anthropic. It's instinct. The coolest technology should make the most money.

But history proves repeatedly: in technology paradigm shifts, the most money flows to companies providing "connective tissue" between the old and new paradigms.

Internet era: Cisco (routers) made more money than most content companies.

Mobile era: Qualcomm (chips) and Apple (OS + distribution) made more money than most app developers.

Cloud era: AWS (infrastructure) made more money than most SaaS companies.

The AI agent era follows the same pattern. Systemic profits will accrue not to AI model builders, but to infrastructure that enables AI agents to operate in the real world. Payment pipes (Stripe/Coinbase/x402). Network pipes (Cloudflare). Commerce pipes (Shopify/UCP). Enterprise pipes (Salesforce/Agentforce). Data pipes (Palantir/Databricks).

If you're thinking "which AI company should I invest in," try a different question: "No matter which AI wins, whose shovels will definitely be used?"

---

Sources: Coinbase, Stripe, Shopify, Google, Salesforce (FY26 Q3 earnings, Dreamforce 2025), Cloudflare (Q4 2025 earnings), Gartner, Chartbeat, Ahrefs, Raptive, Semrush, Press Gazette, IDC, Deloitte, Precedence Research, CIO.com

Disclaimer: This article does not constitute investment advice. All companies and products mentioned are for analytical purposes only.