Anthropic’s roughly $1.5 trillion fully diluted IPO valuation is a bet that a company selling intelligence can keep enough of the value that intelligence creates. I am not so sure.
As I sat looking at my super-red portfolio post the IPO wondering how to tell my wife that our networth is significantly impacted, I was amused as to how the market is reacting as if Anthropic is the only company which can capture the maximum value from the AI boom. The true winners will be a chip to applications mega corporations, Google and Microsoft are way further along than a frontier lab. Except they are not priced at the same multiple.
AI Agents have made one interesting change, the boundary between products and services is collapsing. It’s all about outcomes, customer wants a claim settled, a delivery completed or a machine kept running. Companies that used to occupy different markets begin competing for the same outcome.
The latest All-In discussion is where this clicked. describes companies extending from applications all the way down to infrastructure, silicon and power. His point about free cash flow is just as useful: old labels such as software, hardware and services tell us less about what a business can become.
Every company is becoming an AI company, because AI is becoming part of how the work gets done.
All strategy follows from there. Decide which work to own, which capabilities to buy, and which dependencies could eventually determine your margin. Then invest where greater control produces a better business. Is that it, CEO roles anyone?
So, how does build their strategy in the AI Era?
I would build the strategy around one assumption: agents are inevitable. The pace will vary by task, and human authority will remain essential where consequences are high. But the incentive is straightforward: once a system can complete a job reliably at a lower total cost, a competitor has a reason to use it.
Consider accounting. A software company sells tools that help prepare the accounts. An outsourcing company sells the work of preparing them. An agent that prepares the accounts begins competing with both.
The customer still cares about accuracy, accountability and price. Whether the supplier calls itself a product company or a services business is irrelevant.
This also changes who chooses the inputs. An agent handling a purchase can use one model to read specifications, another to assess a contract and existing software to place the order. The business controlling that workflow can test suppliers, route work and decide when expensive intelligence is worth paying for.
That makes the agent commercially important even when it does not own the underlying model.
AI assistants like Muse receives the customer’s intent before a particular application is selected. For Anthropic, moving from models into agents and applications offers a way to get closer to that decision.
For an incumbent, the uncomfortable question is whether its product remains the place where customers start, or becomes something another company’s agent calls when required. But the key thing is..
We are still very early..
If you go by LinkedIn, every enterprise has deployed an army of agents, retired half its software stack and turned the CFO into a prompt engineer. Actual adoption numbers tell a less glamorous story.
Ramp’s November 2026 AI Index finds that 56.8% of firms in its 85,000-plus-business customer sample paid for AI. Sounds like we are further along. But among firms paying for AI, the median spend is $16.32 per employee per month. The top 1% of AI spenders, by contrast, have a median of about $10,199 per employee. Ramp says just 1% of firms account for roughly 80% of the revenues it observes going to OpenAI and Anthropic.
Source: Ramp Economics Lab, November 11, 2026. These are medians among AI-paying firms in Ramp’s customer base, not national averages or prices of comparable products.
Even the agent numbers need a reality check. Deloitte’s August 2026 survey found that only 15% of surveyed organizations had scaled orchestrated, cross-functional multi-agent systems. Just 5% said their business processes were highly prepared for agentic AI. No wonder: 72% cited a lack of unified, accessible data, 70% raised trust and governance concerns, and 67% said integration was too complex or expensive.
Source: Deloitte, AI Agents Are Only the Beginning, August 12, 2026. The 15% and 5% measure different things in the same research, not sequential adoption stages.
So yes, agents are inevitable. But we are very far from a world in which most companies run their core operations through them. Right now, a lot of the enterprise AI industry is still selling access, experiments and productivity anecdotes.
So I am bullish on the AI opportunity without being automatically bullish on every AI stock. A recent Bain analysis estimates $4.7 trillion in corporate profits at stake through 2035, with 71% of sectors structurally reshaped. More interestingly, roughly three-quarters of that opportunity and disruption comes from innovation and market-share shifts, not simply doing today’s work more cheaply. That’s a much bigger ambition than deploying Copilot and asking HR to report hours saved. A bank can rethink how a loan gets originated, an airline how a disrupted journey gets fixed, and a manufacturer how a fleet of machines stays productive. I made a similar argument in my guide to deploying sovereign agents: the economic unit is the completed transaction, not an impressive chat response.
Follow the money, my brother
Now take that completed transaction and follow the payment backward. Someone owns the customer relationship. Someone operates the agent. Someone supplies the model. Someone hosts the model. Someone makes the chips. And someone has to provide the power, cooling and land. Every layer wants to extract some of the value.
Source: Apoorv Agrawal, The Economics of Generative AI: Two Years Later, April 2026.
That is not a prediction that Nvidia will own 79% of AI forever. It is a snapshot of the buildout phase, when everybody is buying the machinery before the applications have matured. In fact, Agrawal argues that value capture will eventually move upward, as it did in earlier computing cycles. Pretty sure it will.
Think of the game each company is playing. Nvidia would like its chips to remain indispensable while extending into software and serving infrastructure. Hence the push into open models, HuggingFace acquisition etc. Hyperscalers would prefer more choice in chips and more value in their cloud platforms. Model labs need cheaper inference, better distribution and customers who cannot easily switch. Application companies want model providers competing for their workloads, while ensuring an agent doesn’t replace their relationship with the customer.
Everyone wants the layers they buy to become commodities. Nobody wants that to happen to the layer they sell.
And companies are already making moves accordingly. Microsoft launched its own MAI Code 2 and Code 2 Flash models in November 2026, specifically optimized for the price and performance of everyday coding work. It also has Maia 200 silicon, Azure, GitHub, Microsoft 365 and enterprise distribution. Yet it still offers partner models. This is not indecision, it’s a way to own more choices about cost and performance.
Google has TPUs, data centres, Gemini, Cloud, Workspace and Search. Anthropic, meanwhile, is trying to move both ways from the model layer: deeper into enterprise applications and agents, and downward into dedicated compute with the flexibility to run on Nvidia GPUs, AMD GPUs, Google TPUs and AWS Trainium. OpenAI is doing its own version through chips, ChatGPT apps and commerce.
It’s less a tidy market with ten companies occupying six layers and more a strategic bar fight. Your biggest customer can become your competitor. Your biggest supplier can fund your competitor. And the strongest player may not be the one with the best individual component, but the one that can optimize the entire outcome without giving away the margin.
Anthropic is the perfect case study. Also the perfect warning.
Anthropic’s final IPO filing reported $33.24 billion of revenue for the first nine months of 2026. Subtract its $4.81 billion reported cost of revenue and you get an impressive 85.5% gross margin. Sounds like an extraordinary software business.
But the same accounts show $19.24 billion in research and development, $6.62 billion in sales and marketing, and a $1.26 billion operating loss. Compute used to train the next models is part of R&D, not all in the reported cost of serving today’s customer. Cloud-marketplace and channel-partner fees are a meaningful distribution cost.
Most revealing: more than 80% of Anthropic’s Q3 revenue came from models released in the preceding six months. Good product velocity, sure. But how much of the R&D bill is growing new markets, and how much is simply the recurring price of not falling behind? The filing doesn’t tell us, and investors ought to care.
Anthropic has also disclosed approximately $900.9 billion of compute and infrastructure commitments spread over many years, including around $70.7 billion due within the following twelve months. Those commitments are not a single year’s cash burn or a $901 billion debt balance. They do, however, raise the stakes on utilization, costs and the lifespan of the technology being bought.
At the same time, the enterprise opportunity remains enormous. Anthropic has more than 11,900 customers spending at least $100,000 annually, and more than half of the Fortune 500 are in that group. Even so, its average revenue from those Fortune 500 customers is less than 0.1% of their total expenses. There is a great deal of work left to automate. The question is who will get paid for it.
The pressure doesn’t stop with technology. Gary Marcus’s skeptical read of the IPO is worth considering here: spectacular early growth can make it too easy to extend an exponential line for another decade. A reader commenting on Damodaran’s Anthropic valuation makes a pointed argument: if open models keep catching up and prices keep falling, hardware and distribution may capture more of the profits.
Recursive self-improvement adds a twist. Claude is already helping Anthropic build future Claude systems, and AI-assisted engineering is helping model companies optimize chips and code. That’s a real advantage. It is also a capability competitors can improve at. Faster model development could strengthen the leaders, or make model superiority expire even faster. I’ve written before about smaller open models changing the economics of local AI, and I think enterprises will increasingly use different models for different jobs.
Aswath Damodaran’s pre-IPO valuation came to roughly $1.475 trillion in equity value. But his assumptions included revenue approaching $1 trillion by 2036, a long-run 27.5% operating margin, and considerable reinvestment. That is not an easy outcome. It is a spectacularly successful one already required to justify a valuation around the IPO price.
So yes, I can see Anthropic becoming a huge company. I can also see Google, Microsoft, an agent platform, a chip maker, and thousands of well-positioned businesses extracting value from that same growth. The fact that the market grows does not tell you how its profits will be divided.
What I’d ask my board to approve
The wrong takeaway from all this is that every bank, airline, telecom operator and manufacturer needs a frontier lab and its own chip. That’s a very expensive way to learn that not every layer should be yours.
I would start with the outcome. Which customer job could we deliver so much better that we’d gain market share, increase pricing power or change our cost structure? Then I would map every capability required to deliver it, from the application and agent back through the model and compute. Which of those capabilities is a strategic control point, and which is just a procurement decision?
A bank may need to control its customer journey, data, risk evaluation and agent permissions, not train a frontier model. A telecom operator might have a compelling reason to control identity, network-to-edge orchestration and deployment at physical sites, while renting much of the model intelligence. A robotics company might find the expensive, hard-to-copy part is its fleet data, safety envelope and validated operating workflows. That is closely related to my earlier skepticism about general LLMs controlling robots directly: owning a physical outcome is not the same thing as owning every model inside it.
Only after that would I decide whether to buy, partner, co-design or build. The test is financial, not ideological. Does moving into this layer genuinely improve our expected free cash flow after operating costs and reinvestment, at a return above our cost of capital? Or are we simply creating another expensive team because every other CEO seems to have one?
The real strategy test: customer value, control over key dependencies, and returns on the investment required to own them.
This gets even more important when agents erase old product categories. A software vendor might find itself competing with an outsourcer. A machine manufacturer might compete with a fleet-as-a-service provider. A bank might discover that its biggest competitor is the assistant sitting between the customer and every financial product.
At that point, I don’t particularly care whether the company calls itself SaaS, hardware or professional services. Tell me what outcome it controls, what the rest of the stack costs it, what competitors can substitute, and how much cash it can sustainably keep.
Every company is an AI company. All strategy is deciding where you sit in the stack, how far you should extend, and how much of the value you create actually makes it to free cash flow.



