Tomasz Tunguz's Recent LinkedIn Posts

“Hi, Tomasz or Tomasz’s agent.” I’ve started receiving emails that begin this way. A byproduct, I suppose, of having written so much about AI. People now assume my inbox is monitored by robots. Which raises an odd question : what does it mean to write to someone when you expect a machine to answer? Gmail suggests my reply before I’ve thought it. “Sounds good!” “Thanks for sending!” “Let’s circle back next week.” The machine knows what I’d say. Sometimes I click it. Sometimes I wonder if the person on the other end can tell. Every customer support call is now with an AI agent. The voice sounds real. They are infinitely knowledgeable. The responses are fast. Does it matter that it’s not a person? A friend sends voice memos instead of texts now. “So you know it’s actually me,” he said. But how do you know? ElevenLabs can clone a voice from thirty seconds of audio. The ums, the pauses, the little laugh—all reproducible. Does it matter? But maybe the people writing “Hi, Tomasz or Tomasz’s agent” have it right. They’re not being rude. They’re being realistic. They’ve adapted to a world where the answer might come from either side of the curtain, & they’ve decided not to care which. The polite thing now is to assume the robot. The intimate thing is to be surprised when it’s not.

In September 2024, Hurricane Helene flooded Baxter International’s plant in Marion, North Carolina, which produced 60% of the nation’s IV fluids. Within a week, more than 80% of U.S. healthcare organizations reported shortages. One plant, one flood, one week. That disruption made headlines. Most don’t. Eighty-five million packages arrived damaged in the U.S. in 2024, up 30% from the prior year, costing businesses $4 billion. Sean McCarthy saw those failures accumulate during his years at Amazon Shipping, where he was one of the early hires. The investigation process never varied. Query the warehouse management system, often two decades old. Cross-reference the carrier portal. Call the driver, who doesn’t pick up. File a claim: seventeen fields. Four hours pass. Sometimes the problem gets solved. The obstacle was fragmentation. A single shipment can touch 40 to 60 processes across multiple vendors. Connecting them would mean hundreds of bespoke integrations. The project never got funded. Sean partnered with Henry Ou, who led ML teams at Apple and built ranking systems at ByteDance. Together they founded BackOps AI, which deploys AI agents that read emails, click through portals, call drivers, and file claims. When a customer reports a problem, BackOps traces it across every system involved, escalating to a human only when a judgment call is required. We’re leading BackOps’s $26 million Series A. The product works in two stages. Employees record their screens while solving problems; BackOps converts those recordings into automated workflows. Then Relay, the automation engine, runs continuously: filing claims, initiating reshipments, responding to customers. Customers report 93% faster response times and 60% time savings. BackOps files 100% of eligible carrier claims automatically. The platform serves a top global automaker, a leading retailer, major grocery chains, and industrial suppliers. Sean and Henry are targeting a $3.5 billion market growing 13% annually. The bet: AI agents can connect systems that were never designed to talk to each other. So far, the connections hold. If you’d like to learn more, reach out to Sean. Read more from Sean (https://lnkd.in/gNVxdQwp), Theory partner Andy (https://lnkd.in/gHRfs6FU), and Axios (https://lnkd.in/gSrptvD6).

What happens when we run out of GPUs?

In 2025, we predicted that 2026 would be the year agents would earn as much as a person. It’s already happening. In markets where there’s a labor shortage and an urgent need to hire people, we are seeing agents command 75%, 85%, even 100% of a human equivalent salary. This is faster than we were anticipating.

“What happens when a new employee brings their agent to work?” An executive asked this recently. Imagine a few years from now : a student graduates, having trained their own agent through university. It knows everything they’ve learned, every paper, every problem solved. Day one, they bring it to work. It’s like bring your own device circa 2009. The iPhone launched & nobody wanted corporate Blackberries anymore. IT scrambled to adapt. But a rogue phone couldn’t sign contracts. A rogue agent can. Amazon just learned this at scale. $6.3 million in lost orders. 99% order volume drop across North America. Four severity one incidents in one week. Amazon’s AI coding assistant contributed to at least one major production incident. The response : a 90-day safety reset with mandatory two-person review for all code changes. An internal memo admitted what everyone implicitly knows : “Best practices and safeguards around generative AI usage haven’t been fully established yet.” Companies can’t hide behind hallucinations. Utah’s AI Policy Act eliminates the hallucination defense : “It is not an affirmative defense to assert that the GenAI tool made the violative statement or undertook the violative act.” Newer and larger models are smarter and more reliable. But they fail unexpectedly. There is no relationship between size and how failures change over time. AI-generated code creates 70% more issues than human code. The TRUMP AMERICA AI Act would create explicit liability pathways - allowing the US Attorney General, state AGs, & private plaintiffs to sue AI developers for defective design & unreasonably dangerous products. That new hire’s personal agent? The company bears liability for its mistakes. The contracts it signs, the code it deploys, all of it lands on the company. Like a dog or a device, you are responsible for your agent.

In September 2024, Hurricane Helene flooded Baxter International’s plant in Marion, North Carolina, which produced 60% of the nation’s IV fluids. Within a week, more than 80% of U.S. healthcare organizations reported shortages. One plant, one flood, one week. That disruption made headlines. Most don’t. Eighty-five million packages arrived damaged in the U.S. in 2024, up 30% from the prior year, costing businesses $4 billion.

I hate to micromanage & I’ve been micromanaging AI. A few months ago, I’d use Claude for a familiar workflow : capturing notes from a meeting, drafting a follow-up email, updating the CRM, writing the investment memo. Micromanagement at 10x speed. The agent would finish a step, then wait. I’d scan the output, type the next instruction, wait again. Prompt, response, prompt, response. I was the bottleneck in my own system. A year ago, this was necessary. The models couldn’t hold a complex task in their heads. Now they can. But this leverage requires planning. Now I sketch the workflow before I touch the machine. I anticipate the decision branches : what if the company isn’t in the CRM? What if the website is down or the call transcript isn’t available? I flag the gaps before the agent encounters them. This morning’s notebook page : (image) I took a photo & shared it with Claude & walked away. Workflows as images work beautifully. The agents run in the background. The memo sat in my inbox, formatted, sourced, ready to send. Not prompts. Blueprints.

“Hi, Tomasz or Tomasz’s agent.” I’ve started receiving emails that begin this way. A byproduct, I suppose, of having written so much about AI. People now assume my inbox is monitored by robots. Which raises an odd question : what does it mean to write to someone when you expect a machine to answer?

Last week, Cursor launched Composer 2 to over one million daily active users. Within hours, a developer discovered Cursor had built its flagship model on top of Moonshot AI’s Kimi K2.5, a Chinese open-source model. Moonshot AI’s response? “This is the open model ecosystem we love to support.” Cursor’s model is at near parity with state-of-the-art at one-eighth the price.

Kirkland ibuprofen is the same molecule as Advil. Same dosage, same FDA requirements, same therapeutic effect. It costs 80% less. AI has its generic drug moment. DeepSeek V3 matches GPT-5.2 on most benchmarks. It costs 90% less. OpenAI & Anthropic generated $22 billion in 2025. Chinese AI labs generated $1.8 billion. The ratio : 12:1. Pricing explains the gap. Chinese AI API prices collapsed 90% in 2024. US frontier models average $3.38 per million input tokens. Chinese models average $0.48. OpenAI processes roughly 8.6 trillion tokens per day. Chinese labs likely match or exceed this volume. The 12:1 revenue gap isn’t usage. It’s price. Three forces drive Chinese prices down. First, distillation commoditizes capability. Anthropic accused DeepSeek, Minimax & Moonshot AI of conducting “industrial-scale campaigns” to extract knowledge from Claude. OpenAI made similar accusations to Congress. Second, hyperscalers subsidize AI to win cloud customers. Alibaba Cloud cut LLM pricing by up to 97%. Baidu, ByteDance & Tencent spent $1.1B on AI subsidies during Chinese New Year 2026 alone. Third, DeepSeek set the floor. They trained V3 for $6 million versus OpenAI’s $100 million+ for GPT-4, price at $0.14 per million input tokens & hit $220 million ARR with 122 employees. In the US, Chinese models also price at a discount. Together AI charges $1.25 per million input tokens for DeepSeek V3. DeepInfra offers $0.21 per million. DeepSeek’s own API charges $0.14 - 12x less than GPT-5.2. Pharma companies spend billions developing a molecule, then enjoy 20 years of patent protection to recoup R&D costs before generics flood the market. AI follows the same pattern - massive R&D costs upfront, then commoditization. But the timeline is compressed. In pharma, the generic window opens after two decades. In AI, it opens in weeks. DeepSeek V3 costs $0.14 per million tokens. GPT-5.2 costs $1.75. Same capability. Different label. The 90% discount isn’t coming. It’s here. The question : how to protect an asset that takes hundreds of millions to develop when it can be copied in a month?

I hate to micromanage & I’ve been micromanaging AI. A few months ago, I’d use Claude for a familiar workflow : capturing notes from a meeting, drafting a follow-up email, updating the CRM, writing the investment memo. Micromanagement at 10x speed. The agent would finish a step, then wait. I’d scan the output, type the next instruction, wait again. Prompt, response, prompt, response. I was the bottleneck in my own system.

Microsoft's journey with Copilot & NDR metrics in public software companies have a lot in common.

“What happens when a new employee brings their agent to work?” An executive asked this recently. Imagine a few years from now : a student graduates, having trained their own agent through university. It knows everything they’ve learned, every paper, every problem solved. Day one, they bring it to work. It’s like bring your own device circa 2009. The iPhone launched & nobody wanted corporate Blackberries anymore. IT scrambled to adapt.

AI eliminates the marginal hire. Tech job openings are down 45% from the 2022 peak, but up 16% since the start of 2026 - from 227k to 264k. Why the narrative violation? Companies are hiring again, just fewer people than before. A reset to a lower baseline. A team that would have added two engineers to hit next year’s roadmap now ships with the headcount they have. Cursor, Claude Code, Copilot close the gap. The job postings never go live. The offers never extend. Inside most organizations, headcount stays flat. No layoffs. No restructuring announcements. Just fewer new hires than planned. Block slashing 40% of its workforce showed what happens when a company acts on this logic all at once. Jack Dorsey explained : “Intelligence tools we’re creating & using, paired with smaller & flatter teams, are enabling a new way of working which fundamentally changes what it means to build & run a company.” Most companies won’t restructure so dramatically. Until an economic shock, a missed quarter, or pressure from the board forces the question. What AI made possible, AI makes necessary. The restructuring that might have happened gradually over five years happens in one quarter. The seismic shock isn’t coming out of nowhere. It’s building invisibly, one unposted job at a time.

I burned 84 million tokens on February 28th. Researching companies, drafting memos, running agents. That’s running Kimi K2.5, a serverless model via API. At Claude or OpenAI rates — roughly $9 per million tokens blended — equivalent usage would cost $756 for a single day’s work. My peak days hit 80 million tokens. My average days run 20 million. Cloud inference at frontier-model pricing adds up fast.

For every dollar hyperscalers earn from AI today, they’re spending twelve dollars to build more capacity. That’s the bet embedded in $575 billion of capital expenditure this year. How fast does AI revenue need to grow to pay back this data center mortgage?

Last week, Cursor launched Composer 2 to over one million daily active users. Within hours, a developer discovered Cursor had built its flagship model on top of Moonshot AI’s Kimi K2.5, a Chinese open-source model. Moonshot AI’s response? “This is the open model ecosystem we love to support.” Cursor’s model is at near parity with state-of-the-art at one-eighth the price. It’s also no coincidence the editor powering Cursor is open-source, VS Code. $50B in market cap on open-source foundations. Open source empowers startups to compete with incumbents. It’s not easy to replicate Cursor’s innovation on US models. American open-source frontier models average 8 months old. Chinese open-source models average 7 weeks. That’s a 5x age gap. Cursor chose Kimi K2.5 (8 weeks old) over GPT-OSS (8 months old) for good reason : in AI, eight months is three generations of models. Meta, formerly America’s open-source champion with Llama, pivoted to closed-source development in 2025. Chinese open-source models grew from 1.2% of global AI usage in late 2024 to nearly 30% by the end of 2025. Qwen overtook Llama in cumulative downloads by October 2025, reaching 700 million downloads on Hugging Face. But commercializing Chinese models in the US carries risks : NIST found Chinese models 12x more susceptible to agent hijacking attacks, & companies like Microsoft & News Corp have banned their use entirely. Many government agencies have followed suit. Meanwhile, the American open-source response is taking shape. NVIDIA announced a $26 billion commitment over five years to open-source AI through its Nemotron Coalition. Google, OpenAI, & the Allen Institute are building alternatives. OLMo 3 matches Qwen 3 on math benchmarks with 6x less training data. Cursor’s choice wasn’t ideological. It was practical. When the best open-source option is Chinese, that’s what a $50 billion company will use. Open source is how startups compete with giants. The next Cursor will be built on the best open-source foundation available. The question is whether that foundation will be American.

For every dollar hyperscalers earn from AI today, they’re spending twelve dollars to build more capacity. That’s the bet embedded in $575 billion of capital expenditure this year. How fast does AI revenue need to grow to pay back this data center mortgage? From 2020 to 2024, hyperscalers issued an average of $20 billion in bonds annually. In 2025, that jumped to $96 billion. In 2026, it will reach $159 billion. Morgan Stanley projects $1.5 trillion over the next few years. Amazon, Microsoft, Alphabet, Meta, & Oracle will spend 90% of their operating cash flow on AI data centers in 2026, up from a historical average of 40%. Alphabet issued a century bond, the first by a tech company since Motorola in 1997. The debt matures in 2126. Who knows what AI will look like then, or whether Alphabet will exist to repay it. What assumptions justify this borrowing? The depreciation schedules encode the bet. Most hyperscalers depreciate AI infrastructure over five years. At 60% gross margins & 5% borrowing costs, a 5-year payback on $431B in AI capex requires $180B in annual revenue. Current AI revenue is $35 billion. They’re underwriting 5x growth in five years. Nvidia’s stated goal is to release new GPU architectures every twelve months, which will compress depreciation cycles. If chips become obsolete in three years rather than five, the required annual revenue jumps to $276B, 7.9x current levels. As Michael Mauboussin writes, there’s information in prices. The depreciation schedules tell us what hyperscalers believe : AI revenue will grow 5x within five years. The debt markets are betting alongside them.

I burned 84 million tokens on February 28th. Researching companies, drafting memos, running agents. That’s running Kimi K2.5, a serverless model via API. At Claude or OpenAI rates — roughly $9 per million tokens blended — equivalent usage would cost $756 for a single day’s work. My peak days hit 80 million tokens. My average days run 20 million. Cloud inference at frontier-model pricing adds up fast. This week, Alibaba released Qwen3.5-9B, an open-source model that matches Claude Opus 4.1 from December 2025. It runs locally on 12GB of RAM. Three months ago, this capability required a data center. Now it requires a power outlet. A $5,000 laptop — a MacBook Pro with enough memory to run Qwen locally — pays for itself after 556 million tokens. At my usage rate, that’s about a month. At 20 million tokens per day, it’s four weeks. After payback, the marginal cost drops to electricity. It isn’t an intelligence compromise. Reasoning, coding, agentic workflows, document processing, instruction following : the 9B model matches December’s frontier across the board. What changes when frontier intelligence runs locally? Everything I send to cloud APIs today — drafting emails, researching companies, writing code, analyzing documents — stays on my machine. No API logs. No third-party retention. No outages. No rate limits. The tradeoff is parallelization. Cloud APIs handle thousands of concurrent requests. A laptop runs one inference at a time. For simple tasks — summarization, drafting, Q&A — that’s fine. Queue them up. Let them run overnight. For complex agentic workflows that spawn dozens of parallel threads, local inference may not be worth the wait. The economics favor depth over breadth : fewer tasks, run longer, run cheaper. Three months from data center to laptop. The buy-vs-rent math just changed.

“Since last November, 100% of my code has been written by Claude Code. I have not manually edited a single line, shipping 10 to 30 PRs per day.” Boris Cherny, creator of Claude Code, ships 20-30 pull requests per day. Major code changes, not typo fixes. He runs five parallel AI instances, each on a separate branch. Compare that to a traditional engineer : 3 PRs per week. Cherny isn’t 10% more productive. He’s 30x more productive.

AI eliminates the marginal hire. Tech job openings are down 45% from the 2022 peak, but up 16% since the start of 2026 - from 227k to 264k. Why the narrative violation?

In 2025, we predicted that 2026 would be the year agents would earn as much as a person. It’s already happening. In markets where there’s a labor shortage and an urgent need to hire people, we are seeing agents command 75%, 85%, even 100% of a human equivalent salary. This is faster than we were anticipating. The first-order benefit is completing the work. But there are second-order benefits that are now starting to appear. Training agents is significantly faster since all materials can be presented at once & in parallel to the AI. Agents typically require less management burden. They can work 24 hours faster or slower as the team needs. Capacity scales as a function of willingness to spend on inference. Then, a third-order benefit : significantly lower tax burden. Robotic workers are not taxed to the same extent as humans. No FICA. No state unemployment insurance. No benefits. At least a 25-30% cost reduction for the same salary. Plus agent software cost is tax-deductible up to $2.56m. In other categories where AI is augmenting existing workers, the sale is different. Here, the sale captures the marginal hire rather than a big swath of the team. In both conversations, usage tends to surge because of the effectiveness of the systems, much faster than both the vendor and the buyer anticipate. At that point, the business often pauses because a strategic review of organizational design needs to take place. The market rewards this shift. Goldman Sachs found that low-labor-cost stocks outperformed high-labor-cost stocks by 8 percentage points in 2025. Labor’s share of GDP hit a record low of 53.8% in Q3 2025. The implication : every dollar shifted from labor to software improves margins & stock performance. Across the S&P 500, labor costs represent about 12% of revenues on average. Software costs sit around 1-3%. As agents absorb labor, that ratio inverts. Labor shrinks. Software expands. The total addressable market for software grows at labor’s expense, while profitability grows. In the short term, this means no pricing competition on a per-agent basis. Vendors aren’t racing to the bottom ; they can price at par to a person.

“Since last November, 100% of my code has been written by Claude Code. I have not manually edited a single line, shipping 10 to 30 PRs per day.” Boris Cherny, creator of Claude Code, ships 20-30 pull requests per day. Major code changes, not typo fixes. He runs five parallel AI instances, each on a separate branch. Compare that to a traditional engineer : 3 PRs per week. Cherny isn’t 10% more productive. He’s 30x more productive. That productivity gap compounds at the company level. Anthropic generates ~$5 million per employee. Cursor, $3.3 million. Midjourney, $2 million. Traditional SaaS considers $200-300k strong. A 10-20x difference. One explanation : communication overhead. The math follows Metcalfe’s Law. Each new team member adds n-1 new connections. Coordination drag doesn’t grow linearly. It explodes. Now consider what AI does to this equation. A traditional 150-person organization runs four layers deep. The org chart creates 11,175 potential communication channels. Meetings multiply. Alignment decays. An AI-enabled team producing equivalent output might need 30 people. Communication channels drop to 435. A 96% reduction. This is one reason AI-native startups are pulling ahead, and why building AI companies feels fun. The advantage comes from organizational structure. Fewer humans, fewer channels, faster iteration, compounding speed. R&D adopts this fastest. AI writes the code. Human communication becomes the bottleneck. The span of control debate shifts from “how many people can one manager oversee?” to “how many AI agents can one human orchestrate?” Small teams have always paid less coordination tax. AI cuts it further.

I set up a race today between two robots. My Mac on the left vs Claude Code on the right. Both tasked with building a payment app on Stripe’s new Tempo blockchain. Same prompts, same task, side by side. Opus 4.5 is about 20% smarter than Qwen 35B on benchmarks. And it’s likely 50x larger. The hare should have won. It didn’t.

Wiley Jones & Arnav Sahu thrilled to be partners!