Burry Was Right About the Chips. He Didn't Know About the Software.

THE NUMBER: 10x (and 0). That’s the efficiency gain of NVIDIA’s next-gen Vera Rubin chip over current hardware, and the book value of every GPU it replaces.

Last night NVIDIA (NASDAQ: NVDA) reported Q4 earnings: $68.1 billion in revenue, up 73% year over year, $62.3 billion from data centers alone, and guided Q1 to $78 billion (Street expected $73 billion). Jensen Huang declared “the agentic AI inflection point has arrived” and coined a new line: “Compute equals revenues.” Every newsletter tomorrow morning will lead with the beat. They’ll miss the real story.

Vera Rubin samples shipped to customers this week. The next-gen rack delivers 5x the inference performance and 10x lower cost per token than Blackwell. It ships H2 2026. When your replacement chip is 10x more efficient, what is the economic useful life of the chip it replaces? Michael Burry has been arguing that hyperscalers are understating GPU depreciation by $176 billion through 2028. Vera Rubin just made his case for him.

Now apply that same logic to software. This week, Anthropic launched Cowork enterprise plugins (an app store for AI skills) and Claude Code’s claim that it can read and modernize COBOL triggered a 13% drop in IBM (NYSE: IBM), wiping $31 billion in market cap. IBM fired back that translation isn’t modernization. The market didn’t care. When an AI can read the code that locks customers into your platform, the lock-in premium in your stock price evaporates.

Hardware depreciates faster. Software depreciates faster. And now, Guidde just raised $50 million to train AI agents by watching expert video of tasks, not reading documentation. It’s visual imitation learning: the system watches a human use software, captures DOM changes synced with video frames, and builds a digital world model. Pair that with the Chinese distillation story (16 million stolen Claude conversations used to train competing models) and the pattern is clear: AI systems are getting better at extracting knowledge from everything, including each other. The useful life of proprietary processes is shrinking alongside the useful life of chips.

The through-line this week: depreciation is compressing across the entire stack, from silicon to software to institutional knowledge. The market hasn’t priced it in yet.

NVIDIA Beat Earnings. Vera Rubin Beat Everything Else.

NVIDIA (NASDAQ: NVDA) crushed Q4: $68.1 billion total revenue (+73% YoY), $62.3 billion from data centers (+75%), $1.62 per share versus $1.53 consensus. Q1 guidance: $78 billion, five billion above the Street. Stock up 3% after hours. Eight consecutive beats. The machine keeps running.

The headline number isn’t the story. The story has 1.3 million components and a liquid cooling system.

Vera Rubin NVL72: 72 Rubin GPUs, 36 Vera CPUs, 3.6 exaflops of inference, first 100% liquid-cooled rack in NVIDIA’s history. CNBC got the exclusive: 5x greater inference, 10x lower cost per token than Blackwell. Ships H2 2026.

Here’s what Jensen won’t say but Michael Burry has been screaming: when the new chip is 10x more efficient per watt, what’s the old one worth? Google (NASDAQ: GOOGL) and Microsoft (NASDAQ: MSFT) extended GPU useful life estimates from 3 years to 6 years. Meta (NASDAQ: META) went from 3 to 5.5. Burry estimates this understates depreciation by $176 billion from 2026 to 2028. NVIDIA’s counter: CFO Colette Kress says A100s shipped six years ago still run at full utilization.

“Still running” and “economically competitive” are different conversations. An A100 at full tilt burns 10x more electricity per token than Vera Rubin will. When data center power is the binding constraint (Jensen’s own words), old chips aren’t assets. They’re liabilities with a power bill. Google, running its own TPUs, isn’t on this treadmill. Everyone else is on Jensen’s upgrade cycle, and Vera Rubin just cranked the speed.

What this means: The blast radius here goes well beyond the hyperscalers’ boardrooms. If Burry’s thesis plays out, the stockholders in NVDA’s biggest customers get crushed as write-downs eat free cash flow and compress multiples. Then there’s the VC industry, which has placed massive bets on AI infrastructure companies whose unit economics assume GPU fleets hold value for 5-6 years. They don’t. Not anymore. If you own hyperscaler stock, ask what the real depreciation schedule looks like. If you’re a GP who funded an AI infra startup, ask what your hardware is worth in 18 months. Google’s the only one not sweating, because they built their own silicon. Everyone else is paying tuition on Jensen’s treadmill.

Sources:

Claude Reads COBOL. IBM Lost $31 Billion. Software Depreciation Just Arrived.

The stock market is a discounted cash flow machine. Estimate future cash flows, apply a discount rate, get a price. The embedded assumption for enterprise software: customers don’t leave, switching costs are high, obsolescence happens slowly enough to amortize. This week, Anthropic broke that assumption.

Anthropic launched Cowork enterprise with a plugin architecture (Shelly Palmer called it “the app store for AI skills”). Third-party connectors snap into Claude’s desktop. ServiceNow, Notion, Salesforce. This is the platform play: Anthropic stops being a model company and starts being an ecosystem.

Then the detonation. Claude Code demonstrated it can read, interpret, and work with COBOL. IBM(NYSE: IBM) dropped 13%, erasing $31 billion in market cap. IBM argued reading legacy code isn’t modernizing it. They’re right on the technicality. Wrong on the market signal.

Nobody writes new COBOL. IBM’s lock exists because nobody else can maintain it. Banks, insurers, and government agencies pay enormous sums to keep ancient systems alive because full rewrites cost more and carry catastrophic risk. The moment an AI can read that code, debug it, and start translating it, the calculus changes. You don’t need full modernization overnight. You need the optionto leave. Options reprice lock-in premiums instantly.

This is the Burry thesis applied to software. If GPU depreciation schedules are fiction because new chips make old ones uneconomic, software depreciation is equally fictional when AI can replicate what took human teams years to understand. If you thought your legacy vendor had you locked for 3 years, maybe it’s 3 quarters. For a stock priced on cash flows stretching to 2030, that’s a re-rating.

It won’t stop at COBOL. Claude Code won 85% of matches in LLM Skirmish, where LLMs write code to battle each other in real-time strategy games. GPT 5.2 managed 68%. Gemini collapsed from context rot. The coding gap is widening. Grok, OpenAI, and Google will ship their own Cowork equivalents. When they do, the depreciation clock ticks faster.

The action item: If you’re a CEO, CIO, or CTO, this week is a wake-up call. Pull out the list of every system you’ve deployed and everything on the drawing board. For each one, ask two questions: “Can an AI agent read, understand, and eventually replace what this vendor does for us?” and “Will this company still exist in its current form in five years?” If the answer to the first is yes and the second is uncertain, you need to evaluate whether those capabilities can come in-house. For investors, the exercise is the mirror image: find every company in your portfolio whose valuation depends on switching costs and ask what the terminal value looks like when AI gives customers the option to leave. IBM’s shareholders learned this week what happens when that option appears overnight.

Sources:

Guidde Raised $50M to Teach AI by Watching. The Distillation Wars Just Got a New Front.

Guidde closed a $50 million Series B for what it calls “visual imitation learning.” Instead of training AI agents on documentation or code, you show them a video of an expert performing the task. The system captures DOM changes synced with video frames and builds a “digital world model.” The agent learns by watching, not reading.

This sounds like a niche product story. It’s a new front in the knowledge extraction wars.

Nate’s Substack published a deep dive this month on Chinese AI labs using 16 million stolen Claude conversations to train competing models. That’s distillation: using one AI system’s outputs as training data for another. Anthropic called it piracy, not espionage. The distinction matters. Espionage implies nation-state actors. Piracy implies everyone’s doing it.

Visual imitation learning is distillation’s cousin. Instead of capturing text outputs, you capture behavior. “I’m not scraping your content. I’m using your system and noticing what content is there.” The legal distinction is real. The practical distinction is zero.

It’s spreading. OpenClaw users are deploying Scrapling, an open-source tool that bypasses anti-bot systems like Cloudflare Turnstile. Over 200,000 downloads. Cloudflare (NYSE: NET) CTO Dane Knecht: “We make changes, and then they make changes.” A memecoin ($Scrapling) pumped and dumped inside five hours. Welcome to 2026.

Text distillation. Video imitation. Live web scraping. Each method bypasses a different protection layer. The moat you built around your proprietary process assumed human-speed learning. AI doesn’t learn at human speed.

Why it matters: Your competitive advantage used to be “it takes years to train people to do this.” Visual imitation learning compresses that to a video. Ask your team: “What do we know how to do that an AI agent could learn by watching us do it?” If the answer is “most of it,” your differentiation strategy needs to change before your competitors figure out the same thing.

Sources:

Tracking

CJNG Drug Cartel Used AI-Powered Drones and TikTok Recruitment (WIRED) — El Mencho killed, 54% of detected cartel recruitment accounts used AI, 260 drone attacks in H1 2023. The future of conflict is automated.
AIs Recommend Nuclear Strikes in 95% of Wargame Simulations (New Scientist) — King’s College London study: no model ever surrendered, 86% had unintended escalation. Worth reading before anyone puts an LLM near a defense system.
Gemini Can Now Book You an Uber on Your Phone (WIRED) — Google’s Task Automation ships with Galaxy S26 on March 11. The decade-old promise of “Siri, get me an Uber” finally works. MCP integrations, on-device agent navigation, and the hard problem of authentication for agents.
Agent Experience (AX) Is the New UX (CIO) — Netlify CEO coins “Agent Experience.” If your product doesn’t work well with AI agents, agents will route around you. The enterprise version of “design for mobile or die.”

The Bottom Line

Depreciation is compressing across the entire technology stack, and the market is just starting to figure out what that means for valuations.

The useful life of your hardware just shortened by half. Vera Rubin delivers 10x more efficiency than what shipped 18 months ago. Every GPU fleet purchased in 2025 is now on an accelerated depreciation curve the accounting hasn’t caught up with. Google’s the only hyperscaler not holding the bag, because they built their own silicon. Everyone else is financing Jensen Huang’s roadmap. Ask your CFO what the real write-down looks like.

The useful life of your software lock-in is next. IBM lost $31 billion in a day because an AI read COBOL. The market didn’t care that Claude can’t fully modernize those systems yet. It cared that the option to leave now exists. Every enterprise vendor whose stock price depends on high switching costs should be looking at that chart and wondering who’s next.

The useful life of your proprietary knowledge is shrinking fastest of all. Visual imitation, distillation, web scraping: AI systems are learning from everything, including each other. The moat you built on “it takes years to train people to do this” is a sandcastle at high tide. If you haven’t audited what an AI agent could learn by watching your workflows, you’re already behind.

The 2027 winners won’t be the companies with the best technology. They’ll be the ones who understood that everything depreciates faster now and positioned accordingly.

Only the paranoid survive.” — Andy Grove

Key People & Companies

Name	Role	Company	Link
Jensen Huang	CEO	NVIDIA	X
Colette Kress	CFO	NVIDIA	LinkedIn
Michael Burry	Founder	Scion Asset Management	X
Dario Amodei	CEO	Anthropic	X
Shelly Palmer	CEO	The Palmer Group	X
Dane Knecht	CTO	Cloudflare	LinkedIn
Harry DeMott	Author	CO/AI	LinkedIn

Sources

On Repeat: Obsolete by Fear Factory — because when everything you built depreciates faster than you planned, the only question is whether you saw it coming or got run over by it.

Compiled from 18 sources across CNBC, Fortune, Axios, WIRED, VentureBeat, CIO, and independent research. Cross-referenced with thematic analysis and edited by Harry DeMott and CO/AI’s team with 30+ years of executive technology leadership.

Burry Was Right About the Chips. He Didn’t Know About the Software.

THE NUMBER: 10x (and 0). That’s the efficiency gain of NVIDIA’s next-gen Vera Rubin chip over current hardware, and the book value of every GPU it replaces.

NVIDIA Beat Earnings. Vera Rubin Beat Everything Else.

Claude Reads COBOL. IBM Lost $31 Billion. Software Depreciation Just Arrived.

Guidde Raised $50M to Teach AI by Watching. The Distillation Wars Just Got a New Front.

Tracking

The Bottom Line

Key People & Companies

Sources

Past Briefings

OpenAI Deleted ‘Safely.’ NVIDIA Reports. Karpathy Is Still Learning

Altman lied about a handshake on camera. CrowdStrike fell 8%. Google just killed the $3,000 photo shoot.

We’re Building the Agentic Web Faster Than We’re Protecting It