NextFin News - Anthropic, the high-profile artificial intelligence startup backed by tech giants like Amazon and Google, has significantly lowered its long-term profit margin outlook. According to reports from The Information on January 22, 2026, the company adjusted its gross margin projections as the astronomical costs of training and operating its large language models (LLMs) continue to climb. While Anthropic has seen its revenue skyrocket over the past year, the internal revision suggests that the path to sustainable profitability is becoming increasingly narrow due to the rising "AI operating tax."
The news comes as Anthropic deepens its reliance on specialized hardware to maintain its competitive edge. Just this week, Broadcom CEO Hock Tan confirmed that Anthropic is the previously unnamed "fourth customer" for its custom AI chips, placing a massive $11 billion order for delivery in late 2026. This follows an initial $10 billion commitment, bringing Anthropic’s total hardware backlog with Broadcom to over $21 billion. Despite these aggressive investments aimed at long-term efficiency, the immediate reality for Anthropic is a margin squeeze driven by the sheer scale of compute power required to run its Claude 3.5 and Claude 4 series models.
The downward revision of margin expectations is a sobering signal for the venture capital community and the broader AI sector. For years, the software-as-a-service (SaaS) model enjoyed gross margins of 70% to 80% because the marginal cost of serving an additional customer was near zero. AI, however, breaks this mold. Every query processed by an LLM incurs a tangible cost in electricity, cooling, and GPU/TPU utilization. According to industry analysts, Anthropic’s move reflects a structural shift where the cost of goods sold (COGS) for AI companies is fundamentally higher than traditional software, potentially capping gross margins in the 50% to 60% range for the foreseeable future.
This margin pressure is exacerbated by the current competitive landscape. To keep pace with OpenAI and Google, Anthropic must continuously train larger models on more expensive datasets. The $11 billion order with Broadcom indicates that Anthropic is attempting to move away from off-the-shelf GPUs toward custom silicon (ASICs) to lower its per-token costs. However, the capital expenditure (CapEx) required for such a transition is immense. While custom chips may eventually offer better performance-per-watt, the upfront R&D and manufacturing costs are weighing heavily on the company’s current financial outlook.
Furthermore, the concentration of the AI supply chain adds another layer of risk. Broadcom’s recent financial disclosures show that its $73 billion backlog is concentrated among just five major customers, including Anthropic. This mutual dependency means that any delay in chip delivery or a shift in U.S. trade policy under U.S. President Trump could have outsized impacts on Anthropic’s ability to scale. As U.S. President Trump emphasizes domestic manufacturing and potential tariffs on high-tech components, the cost of the hardware underlying Anthropic’s infrastructure remains volatile.
Looking ahead, the industry is likely to see a "flight to efficiency." If Anthropic cannot maintain high margins through raw model performance, it will be forced to innovate in model distillation and architectural efficiency. We expect to see more "small language models" (SLMs) that offer 90% of the capability at 10% of the operating cost. For Anthropic, the challenge in 2026 will be proving to investors that its skyrocketing revenue can eventually decouple from its skyrocketing compute bills. Without a breakthrough in how AI is served at scale, the era of hyper-profitable software may be giving way to a more capital-intensive, lower-margin industrial reality for the age of intelligence.
Explore more exclusive insights at nextfin.ai.
