Breaking: Anthropic IPO Math Hinges on GPU Payback and Compute Control
A single high-end graphics card serves limited users, creating a massive gap between revenue and valuation as Anthropic prepares to list.
By Julia Wang ·
Anthropic is preparing for an initial public offering that values the company between 2.5 trillion dollars and 3 trillion dollars. The company projects this valuation despite carrying compute commitments to partners like Amazon and Google totaling up to 900 billion dollars. Current financial reports indicate Anthropic is running a deficit of about 500 billion dollars while generating less than 20 billion dollars in annual revenue. The central question for investors is how a company losing 500 billion dollars annually can command a valuation in the trillions. The answer lies in the control of compute infrastructure and the economics of serving AI models to users.
The foundation of this valuation rests on the hardware required to run the models. A single Nvidia RTX Pro 6000 graphics card is the unit of analysis for this cost structure. This specific card 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. It also excludes the data center facility costs or the electricity required to keep the systems running. For a company seeking to maximize returns per dollar of hardware, the efficiency of this single card determines the viability of the entire business model.
Under constant use twenty four hours a day, one RTX Pro 6000 card can hold about 15 to 20 light users. These users run light compute at a speed of about 60 to 70 tokens per second. Alternatively, the same card can serve 1 to 2 heavy coding users running Claude Code at about 40 to 50 tokens per second. The heavy users pay 200 dollars a month for the max plan. The light users pay about 20 dollars a month on average for a base plan. This disparity in pricing and usage intensity creates a specific payback period for the hardware investment.
When buying the cards directly from Nvidia at a wholesale price of 3,000 to 4,000 dollars, the GPU alone takes about 10 months to pay back for a heavy user paying 200 dollars a month. The breakeven point is roughly the same for the 15 to 20 light users paying 20 dollars a month. However, most light users will not use the chat for more than a few hours a day. This behavior can stretch the bandwidth number by up to 10 times. Despite this potential for higher utilization, the acquisition cost and breakeven period remain significant barriers for the company.
Nvidia, the primary supplier of these GPUs, understands this bottleneck well. The company drives demand up and lowers supply while allocating all its supply to large players like Anthropic, Grok, and OpenAI. This strategy leaves the consumer market fitting the bill for the infrastructure. The compute and GPU bottleneck is a real problem for AI companies, and Nvidia leverages this position to maintain market control. The high cost of entry ensures that only well capitalized firms can sustain the necessary infrastructure.
The control over distribution has become a key battleground in the AI market. Nvidia agreed to acquire Hugging Face, the platform where most open AI models are distributed, for 12.9 billion dollars. Hugging Face makes about 150 to 200 million dollars a year in its peak years, making the acquisition price about 70 times its revenue. While 12.9 billion dollars is pocket change for Nvidia and its CEO Jensen Huang, the strategic value lies in controlling the distribution of models. The stated reason for the deal was CUDA, the software that runs all these models, but the deeper reason was control over the open source market.
The open source market was previously flooded by advanced Chinese models, including Kimi from Moonshot AI, DeepSeek, Qwen from Alibaba, and models from Xiaomi. These companies were willing to give the models over to the public to increase innovation. Kimi K3 started to outperform the Opus 5 models from Anthropic and 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 platform.
Developers on Discord rallied to stop this censorship and corporate overreach. 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. This move by Nvidia and Anthropic highlights the tension between open distribution and proprietary control. The removal of competitive open source models strengthens the position of closed systems like Claude.
Congress has played a role in shaping this market through legislation. Anthropic demanded that Congress ban Chinese models because of national security, though the real goal was to control the market. Anthropic recently executed a publicity stunt by circulating 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. The panic worked, and Congress passed a bill that now bans any government and its affiliates from using any Chinese models.
Anthropic is also trying to pass bills that would stop other companies like Verizon and other networks from launching their own AI models in house by using these advanced open source models. This legislative push complements the hardware control strategy. By limiting who can use open source models, Anthropic protects its revenue streams from internal competition. The combination of GPU scarcity and legislative barriers creates a robust moat around the company.
