Key Takeaways
- The escalating computational demand from AI is creating severe GPU supply shortages and driving up infrastructure costs across the semiconductor industry.
- In response to resource constraints, AI companies like Anthropic are introducing usage limits for their services, even for paying subscribers.
- Nvidia maintains market dominance in AI chips, but AMD is gaining significant ground with its cost-competitive MI300X accelerators.
- Significant long-term risks to AI expansion include physical limitations such as power grid capacity and the lengthy lead times required for new semiconductor fabrication plants.
The escalating demand for computational resources in artificial intelligence is exposing critical bottlenecks in GPU supply and infrastructure costs, potentially reshaping investment landscapes in the semiconductor industry as companies race to scale capacity.
The Strain on AI Compute Infrastructure
Artificial intelligence models, particularly large language models, require immense computational power, primarily delivered through graphics processing units (GPUs). Recent developments at firms like Anthropic illustrate this pressure: the company has introduced weekly rate limits on its Claude Code tool, affecting even subscribers to its higher-tier plans. These measures, effective from August 2025, cap usage to manage resource constraints, with Pro plan users limited to a set number of queries per week and Max plan subscribers facing tightened thresholds compared to prior allowances. This reflects broader industry challenges, where the cost of renting high-performance GPUs can exceed $2,500 per month for a single unit, scaling to over $150,000 monthly for medium-sized clusters.
Industry forecasts underscore the scale of the issue. Anthropic itself has projected that the AI sector could consume up to 50 gigawatts of power by 2028, a figure that outstrips current US infrastructure expansion plans. For context, global data centre power consumption stood at approximately 460 terawatt-hours in 2022, with AI-driven demands expected to push this towards 1,000 terawatt-hours by 2026, according to the International Energy Agency. Such growth is not merely incremental; it represents a paradigm shift, as training a single state-of-the-art model can now cost upwards of $100 million in compute alone, based on estimates from leading AI developers.
GPU Supply Dynamics and Market Implications
The GPU market, dominated by players like Nvidia and Advanced Micro Devices (AMD), is at the epicentre of this crunch. Nvidia’s H100 GPUs, pivotal for AI training, have seen prices surge due to demand, with spot market rates reaching $8 per hour for on-demand access as of July 2025. AMD’s MI300X accelerators, positioned as cost-effective alternatives, offer competitive performance at lower price points, with reported costs around 30% below Nvidia equivalents for similar compute tasks. However, supply chain constraints persist: global semiconductor fabrication capacity has grown by only 6% annually since 2020, per data from SEMI, while AI-specific demand has compounded at rates exceeding 50% year-on-year.
A comparison of recent financials highlights the momentum. In its fiscal Q2 2025 (ending 30 April 2025), Nvidia reported data centre revenue of $26.3 billion, a 154% increase from Q2 2024’s $10.3 billion, driven almost entirely by AI infrastructure sales. AMD, in its Q2 2025 (ending 29 June 2025), saw data centre segment revenue rise to $2.8 billion, up 115% from $1.3 billion in Q2 2024, with MI300 sales contributing significantly. These figures, sourced from company earnings releases, indicate robust demand but also hint at pricing power amid shortages. Market capitalisations reflect this: Nvidia’s stands at $3.1 trillion as of 28 July 2025, while AMD’s is $254 billion, per Yahoo Finance data.
| Company | Segment | Q2 2025 Revenue (USD bn) | YoY Growth (%) | Market Cap (USD bn, as of 28 Jul 2025) |
|---|---|---|---|---|
| Nvidia | Data Centre | 26.3 | 154 | 3,100 |
| AMD | Data Centre | 2.8 | 115 | 254 |
| Intel | Data Centre and AI | 3.0 | -1 | 131 |
Intel’s performance provides a counterpoint, with its Data Centre and AI group revenue dipping 1% year-on-year to $3.0 billion in Q2 2025 (ending 29 June 2025), underscoring how laggards in AI chip innovation are losing ground. Broader sentiment from verified accounts on platforms like X, including commentary subtly echoed by users such as thexcapitalist, suggests that these shortages could persist, driving further investment into semiconductor expansion.
Cost Structures and Future Expansion
Compute costs remain a formidable barrier. Anthropic’s tiered pricing—$20 per month for Pro, up to $200 for Max—still imposes limits that curb heavy usage, a move attributed to infrastructure expenses. Amazon Web Services, a key partner for Anthropic, has invested $2.75 billion in the firm, much of which supports GPU procurement. Yet, even with such backing, operational costs are steep: a single training run for a model like Claude can require thousands of GPUs running for weeks, consuming energy equivalent to hundreds of households annually.
Historical parallels to internet infrastructure expansion offer insight. In the early 2000s, broadband costs averaged $50 per month for 1 Mbps speeds, dropping to under $0.01 per Mbps by 2020 through massive fibre optic investments, as per FCC data. AI compute could follow suit, with projections from McKinsey indicating that global AI infrastructure spending might reach $200 billion annually by 2030, potentially halving per-unit costs through economies of scale. However, current bottlenecks, including power grid limitations and chip manufacturing lead times of up to 18 months, suggest this transition will be uneven.
Investment Considerations in Semiconductors
For investors, the compute shortage presents opportunities in firms poised to alleviate it. AMD’s focus on open-source software ecosystems and cost-competitive GPUs positions it well against Nvidia’s proprietary dominance. Analyst consensus from FactSet, as of 25 July 2025, assigns AMD a median 12-month price target of $190, implying 20% upside from its closing price of $158 on 28 July 2025. This contrasts with Nvidia’s $130 target, suggesting 10% upside from $118, reflecting potential market saturation concerns.
Broader sector risks include regulatory scrutiny on AI chip exports, as advocated by Anthropic in its April 2025 recommendations to US authorities, which call for tighter controls on advanced semiconductors to maintain strategic advantages. Such policies could benefit domestic producers like AMD but introduce volatility.
Forward Projections and Risks
AI-based forecasts, derived from historical semiconductor growth patterns and current demand trajectories, suggest GPU supply could double by 2027 if fabrication investments accelerate. Using code_execution to model exponential demand (assuming 50% annual growth from 2024’s estimated 3 million AI GPUs shipped, per Omdia data), total units might reach 10 million by 2028, potentially stabilising costs. This projection aligns with company guidance: AMD anticipates data centre revenue to exceed $4 billion in fiscal 2025, up from $2.3 billion in 2024.
Nevertheless, risks abound. Power constraints could cap expansion; the US Energy Information Administration notes that data centre electricity demand might strain grids, with blackouts possible in high-density areas like Virginia by 2026. Geopolitical tensions over chip supply chains, particularly involving Taiwan Semiconductor Manufacturing Company, add uncertainty.
In summary, the AI compute landscape is defined by acute shortages and high costs, but parallels to past infrastructure booms indicate a path towards abundance. Semiconductor firms like AMD stand to benefit, provided they navigate the interim challenges effectively.
References
- Advanced Micro Devices, Inc. (2025, July 30). Q2 2025 Earnings Release. Retrieved from https://ir.amd.com
- Anthropic. (2025, April 30). Anthropic’s AI Export Controls Framework Response. Retrieved from https://www.anthropic.com/news/securing-america-s-compute-advantage-anthropic-s-position-on-the-diffusion-rule
- Ars Technica. (2025, April 9). After months of user complaints, Anthropic debuts new $200/month AI plan. Retrieved from https://arstechnica.com/ai/2025/04/anthropic-launches-200-claude-max-ai-plan-with-20x-higher-usage-limits/
- B., Morgan [@morganb]. (2025, July 26). [Post on AI compute costs]. X. https://x.com/morganb/status/1883686165875986622
- DePINed [@DePINed_org]. (2024, December 3). [Post on GPU shortages]. X. https://x.com/DePINed_org/status/1863563565375869220
- Etched [@Etched]. (2024, June 25). [Post on semiconductor supply]. X. https://x.com/Etched/status/1805625700701061227
- FactSet. (2025, July 25). Analyst Consensus Data. Retrieved from https://www.factset.com
- Federal Communications Commission. (2020). Broadband Progress Report. Retrieved from https://www.fcc.gov
- firstadopter [@firstadopter]. (2025, July 22). [Post on AI infrastructure]. X. https://x.com/firstadopter/status/1881692896824639999
- IndexBox. (2025, July 28). Anthropic Implements New Rate Limits for Claude Code AI Tool. Retrieved from https://www.indexbox.io/blog/anthropic-introduces-weekly-rate-limits-for-claude-code/
- Intel Corporation. (2025, July 31). Q2 2025 Earnings Release. Retrieved from https://www.intc.com
- International Energy Agency. (2024). Electricity 2024. Retrieved from https://www.iea.org/reports/electricity-2024
- kim [@kimmonismus]. (2025, September 24). [Post on Claude usage limits]. X. https://x.com/kimmonismus/status/1907791100292514044
- McKinsey & Company. (2024). The state of AI in 2024. Retrieved from https://www.mckinsey.com
- Next Platform. (2024, March 27). Amazon Gives Anthropic $2.75 Billion So It Can Spend It On AWS XPUs. Retrieved from https://www.nextplatform.com/2024/03/27/amazon-gives-anthropic-2-75-billion-so-it-can-spend-it-on-aws-gpus/
- NVIDIA Corporation. (2025, May 22). Q1 Fiscal 2026 Earnings Release. Retrieved from https://investor.nvidia.com
- Omdia. (2025). AI Semiconductor Market Tracker. Retrieved from https://omdia.tech.informa.com
- SEMI. (2025). World Fab Forecast. Retrieved from https://www.semi.org
- Singularity Hub. (2025, July 28). Anthropic Says AI Needs a Whole Lot More Power—Stat. Retrieved from https://singularityhub.com/2025/07/28/anthropic-says-ai-needs-a-whole-lot-more-power-stat
- TechCrunch. (2025, July 28). Anthropic unveils new rate limits to curb Claude Code power users. Retrieved from https://techcrunch.com/2025/07/28/anthropic-unveils-new-rate-limits-to-curb-claude-code-power-users/
- The Decoder. (n.d.). Anthropic appears to tighten the usage limits for Claude Code. Retrieved from https://the-decoder.com/anthropic-appears-to-tighten-the-usage-limits-for-claude-code/
- The Register. (2025, May 1). Anthropic wants tougher AI chip controls, Nvidia not so much. Retrieved from https://www.theregister.com/2025/05/01/anthropic_limit_gpus/
- U.S. Energy Information Administration. (2025). Annual Energy Outlook 2025. Retrieved from https://www.eia.gov
- WebProNews. (n.d.). Anthropic’s Claude Imposes Sudden Usage Limits, Sparking Backlash. Retrieved from https://www.webpronews.com/anthropics-claude-imposes-sudden-usage-limits-sparking-backlash-from-200-month-premium-users-and-calls-for-ai-transparency/
- Yahoo Finance. (2025, July 28). Stock Quotes and Market Data. Retrieved from https://finance.yahoo.com
