The Market Said Sell. Big Tech Spent $10 Billion Anyway.

Three thousand construction workers in El Paso are building the most expensive data center ever attempted. This week, Meta 6x’d the budget mid-pour. $1.5 billion became $10 billion. No new product launched. No revenue milestone hit. Just a quiet admission that production AI eats compute at a scale nobody projected publicly—not even Meta.

The S&P 500 fell for its fifth straight week. NVDA dropped 9%.

The market is pricing the Iran war. The companies are pricing the next decade. One of them is wrong—and the gap between those bets is where Q2’s opportunity lives.

Thursday, March 26 close. Week-over-week vs. prior Friday.

* VIX = the “fear gauge.” Measures expected S&P 500 swings over the next 30 days. Above 25 = real anxiety.

Bottom line: The fear index and the capex budgets are telling opposite stories. One of those stories is wrong. I know which one I’m betting ages better.

1. Meta 6x’s Its Data Center—Mid-Construction

Meta took its El Paso AI data center from $1.5B to $10B. No new product. No revenue trigger. Just a revised cost model for what production AI actually requires. The facility targets 1 GW by 2028. Big Tech collectively: $630B+ in AI capex this year.

What it means: A 567% budget revision mid-build isn’t ambition. It’s an accounting correction. The public market’s demand estimates are behind.

2. Three Frontier Models in 23 Days

GPT-5.4, Gemini 3.1 Ultra, Grok 4.20—all in March. The gap between frontier labs is now weeks, not months. Gemini 3.1 Pro matches GPT-5.4 Pro on leading benchmarks at a third of the price.

What it means: Model capability is commoditizing fast. The moat is shifting to whoever owns the infrastructure, the distribution, and the enterprise data. The model leaderboard isn’t where the money is.

3. Google–Agile Robots: The Data Flywheel Nobody’s Talking About

Google DeepMind partnered with Munich-based Agile Robots—20,000+ industrial systems deployed globally—to pipe Gemini Robotics into real hardware. Robots generate field data; data trains better models; better models go to more robots.

What it means: This is the search flywheel, rebuilt for physical AI. Every deployment compounds Google’s lead in ways a competitor can’t shortcut.

4. Robotics Funding Hits $1.2B in a Single Week

Mind Robotics ($500M), Rhoda AI ($450M), Sunday ($165M), Oxa ($103M). Add SkildAI’s $1.4B from February, and 2026 is pacing $20B+ in robotics VC. New annual record.

What it means: Capital moving this fast isn’t chasing hype. It’s funding an infrastructure cycle—the same playbook cloud followed in 2011.

5. OpenAI Kills Sora—First AI Product to Die on Unit Economics

The Sora video API is dead. Compute costs were “economically irreconcilable” with any price users would pay. Enterprise budgets pivoted immediately to Runway Gen-4, Pika 2.1, and Google Veo 2.

What it means: First real casualty. Impressive demos aren’t businesses. The winners at scale will be the ones who can actually afford to run the models.

6. MCP Hits 97M Installs—The Agent Economy Has Plumbing

The Model Context Protocol—open standard connecting AI agents to tools and data—crossed 97 million installs. Every major AI provider ships MCP-compatible tooling now. It went from experiment to infrastructure in under a year.

What it means: The picks-and-shovels play on AI agents isn’t just chips. It’s the middleware. ServiceNow, Salesforce, Workday—the companies owning that connective layer are capturing real platform value, quietly.

7. Tesla Puts $2B Into xAI. The Governance Math Gets Ugly.

Tesla plans to invest ~$2B in xAI—the AI company controlled by Tesla’s own CEO. Tesla stock: down 18.6% in March, the S&P’s biggest single-name market-cap loss.

What it means: This isn’t an EV bet anymore. It’s a bet on one person’s ability to allocate capital across companies he controls. Different thesis. Much harder to underwrite.

8. Five Straight Red Weeks. Capex Didn’t Blink.

S&P 500 is ~6% off its January peak. Iran pushed oil up ~31% since March 1. Airlines down ~20%. Consumer discretionary down ~12%. VIX averaged 24.3 in March vs. 16.1 in February. And through all of it: $630B in AI capex, Meta’s $10B revision, three frontier models.

What it means: The market is pricing fear. The companies are pricing a decade. That gap—prices down, capex up—is Q2’s defining tension.

Everyone’s watching the VIX. The smarter bet is watching Meta’s construction budget.

Here’s what the fear trade misses. The companies building AI infrastructure aren’t waiting. Not for Iran to resolve. Not for the VIX to settle. Not for the Fed. They ran their internal compute models, and the answer came back: spend more, spend faster, spend now. That’s not bravado. That’s a demand signal from inside the machine.

The dots connect if you let them:

• Meta’s $10B (#1) + three frontier models in 23 days (#2) = inference demand is outrunning every public forecast. Models getting cheaper and faster at the same time means usage explodes. Usage is what justifies the spend.

• Google’s flywheel (#3) + $1.2B in robotics funding (#4) = physical AI just entered its cloud-2011 moment. The question isn’t whether robotics is real. It’s whether you’re in before or after the market prices the infrastructure phase.

• Sora dies (#5) + MCP at 97M (#6) = the sorting has started. Capability without unit economics is a dead end. Platform lock-in wins. This is where the field narrows.

The thesis: The market is using a geopolitical lens—Iran, oil, the Fed—to price companies running on a decade-long compute clock. Those are different timeframes. They’re not in conflict. They’re just not the same conversation. Long-horizon investors who see that difference are getting handed a discount.

We break down which AI stack layer has the best risk-adjusted entry right now—and which is most exposed if Iran escalates—in this week’s deep dive. [Link]

Not recommendations—starting points for your own research.

Idea 1: Vertiv (VRT)—The Toll Road on Every AI Data Center

Why now: Meta’s $10B revision (#1). Every gigawatt-scale data center needs cooling, power conversion, and rack infrastructure. Vertiv builds all of it.

The case: $630B in AI capex flows through Vertiv-category infrastructure. Multi-year backlog visibility, no commodity pricing risk. It’s a toll road, not a chip bet.

The risk: Multiples already price in acceleration. Any Big Tech capex guidance cut compresses Vertiv before revenues slow.

Tripwire: Microsoft or Amazon revising 2026 capex down in Q1 earnings (April–May). That’s the exit signal.

How to research: VRT (NYSE). Compare: Eaton (ETN), Schneider Electric (SU.PA).

Idea 2: ServiceNow (NOW)—The Toll Booth on the Agent Economy

Why now: MCP at 97M installs (#6). Enterprises deploying AI agents at scale need workflow orchestration and audit trails. ServiceNow is already the system of record.

The case: Down ~6% on macro, but the AI adoption story is accelerating. Captures per-seat economics on every agentic deployment without building the models. Toll booth, not toll road.

The risk: Premium multiple assumes high net revenue retention. If enterprises pump the brakes on agentic AI, the compression is sharp.

Tripwire: Net new ACV growth below 20% YoY at Q1 earnings (late April) = enterprise AI is stalling.

How to research: NOW (NYSE). Compare: CRM, WDAY, WCLD ETF.

Idea 3: Japanese Robotics—28x for What the U.S. Prices at 58x

Why now: Google-Agile deal (#3) and $1.2B funding week (#4) validate industrial robotics demand. Japanese companies run >50% of global deployments at half the valuation.

P/E ratio = price ÷ earnings. Lower = cheaper per dollar of profit.

The risk: Yen volatility, domestic headwinds, Chinese entrants at subsidized scale. The discount might be justified.

Tripwire: BYD or Huawei formally entering industrial robotics with sub-cost pricing. Then the gap is structural, not a mispricing.

How to research: FANUC (6954.T), Yaskawa (6506.T). Or BOTZ/ROBO ETFs for robotics without the FX risk.

A living thematic framework—not a buy list. Updated weekly. Conviction = structural thesis strength, not this week’s price action.

Changes this week: Platforms goes to HIGH. Three frontier models in 23 days plus Fortune 500 agentic deployments confirmed at GTC 2026—enterprise adoption is pulling forward, and MEDIUM didn’t reflect that anymore. NOW added to Applications on the agentic orchestration thesis.

Conviction check—Infrastructure: NVDA down 8.6%. So why hold? Because the sell-off was Iran, oil, and VIX—not earnings, not guidance, not a demand miss. Meta revised upward. Capex is $630B and climbing. The price moved. The thesis didn’t. We hold.

What we’re watching: (1) Q1 Big Tech earnings (April–May)—Microsoft and Amazon capex guidance is the Infrastructure layer’s stress test. (2) Iran trajectory through Q2—ceasefire triggers a sharp relief rally in AI names. (3) ServiceNow Q1 (late April)—first hard data on whether agentic deployments are real revenue or still PowerPoints.

Today you learned:

1. The builders aren’t waiting for the market. A 567% budget revision mid-construction isn’t optimism. It’s a demand signal from inside the system.

2. The sorting has started. Sora had the capability. It didn’t have the math. Cost structure and platform lock-in—not benchmarks—decide who survives.

3. Two clocks, one market. The market’s clock runs on Iran, the VIX, the Fed. The companies’ clock runs on a decade of compute. If you can tell them apart, you get the discount.

Now research one.

Most people will skim this and check their portfolio Monday morning. The ones who spend 30 minutes this weekend digging into one idea will be building conviction while everyone else is reacting to headlines.

Reply with which one you’re starting with. I read every response.

Stay curious—and stay qualified.

— R. Lauritsen