Breaking: The RTX Pro 6000 Payback Problem in Anthropic's IPO
A 20,000 dollar card takes ten months to recoup from one heavy user, defining the capital intensity behind the 2.5 trillion dollar valuation.
By Thiago Cruz ·

Anthropic is preparing an initial public offering that projects a valuation between 2.5 trillion dollars and 3 trillion dollars. This figure stands in stark contrast to the company's current financial reality, which includes a deficit of about 500 billion dollars and annual revenue of less than 20 billion dollars. The core question for investors is how a company losing that much money can justify such a massive market value. The answer lies in the physical infrastructure required to run its models, specifically the high cost of graphics processing units that serve as the engine for artificial intelligence workloads.
The economics of serving these models are driven by the hardware required to process tokens. A single Nvidia RTX Pro 6000 graphics card can hold about 15 to 20 light users running light compute at about 60 to 70 tokens per second. Alternatively, the same card can support 1 to 2 heavy coding users running Claude Code at about 40 to 50 tokens per second, assuming constant use 24 hours a day. These numbers represent the maximum theoretical capacity of the hardware under ideal conditions of continuous demand.
The cost to acquire this hardware is substantial. An RTX Pro 6000 retails at 20,000 dollars as of now. This price tag does not include the servers, switches, CPUs and processors needed to run the machines, nor does it cover the data center facility costs or the electricity required to keep them running. For a company looking to scale, the entry barrier for high performance compute is extremely high, forcing reliance on bulk purchasing or long term commitments.
When viewed from the perspective of a heavy user, the payback period for this hardware becomes a critical metric. Each heavy user on the max plan pays 200 dollars a month. Buying the cards from Nvidia direct at a wholesale price of 3,000 to 4,000 dollars, the GPU alone takes about 10 months of a heavy 200 dollar a month user to pay back. This calculation assumes that the card is fully utilized by a single heavy user throughout that period, a scenario that is rare in general consumer markets but common in specialized enterprise environments.
For light users, the math shifts slightly but the timeline remains similar. Each light user pays about 20 dollars a month on average for a base plan. The same 3,000 to 4,000 dollar wholesale GPU cost takes about the same amount of time to pay back when spread across 15 to 20 light users. Most light users will not use the chat for more than a few hours a day, so the bandwidth number can stretch 10 times. However, the breakeven and acquisition cost stays huge, meaning that the hardware must be kept busy to justify its expense.
This compute and GPU bottleneck is a real problem for these companies, and Nvidia knows it. Nvidia drives demand up and lowers supply, giving all its supply to the large players like Anthropic, Grok and OpenAI. The consumer market is left fitting the bill, often paying retail prices that reflect the scarcity engineered by the supplier. This dynamic ensures that the largest companies with the deepest pockets can secure the hardware needed to maintain their lead in the artificial intelligence race.
The strategic importance of this hardware is further highlighted by Nvidia's recent acquisition of Hugging Face for 12.9 billion dollars. Hugging Face is the platform where most open AI models are distributed and makes about 150 to 200 million dollars a year in its peak years. The 12.9 billion dollar price is about 70 times its revenue, a sum that is pocket change for Nvidia and its CEO Jensen Huang. The stated reason for the deal was CUDA, the software that runs all these models, but the deeper reason was control of the distribution of models on the market, especially the open source market.
This control is becoming increasingly necessary as the open source market is flooded by advanced Chinese models. Kimi from Moonshot AI, DeepSeek, Qwen from Alibaba, and models from Xiaomi are being given to the public to increase innovation. Kimi K3 started to outperform the Opus 5 models from Anthropic, and it was given to the public. The internet used these large language models, quantized them down, and made them usable on smaller machines like Mac Studios, A100 cards and RTX Nvidia models. After Nvidia acquired Hugging Face, advanced models were deleted from distribution. Anything over a quant 4, meaning advanced and capable of beating out Claude's models, was removed. The 4 bit, 6 bit, 8 bit and 16 bit full sized models were banned from the Hugging Face platform.
Developers on Discord rallied to stop the censorship and corporate overreach, and they opened Pirate Face, a place where you can upload and download all models, but not before the large 400 gigabyte models at quant 3 and higher were scrubbed from the network. Anthropic demanded that Congress ban Chinese models because of national security, when the real goal was to control the market. Anthropic recently did a publicity stunt: it released and circulated media that AI will take over the world when one of its disgruntled employees, Jacob Coxon, quit amidst his disappointment in AI. The goal was to get the government to pass bills banning models that were a huge threat to Anthropic's bottom line.






