AI Bubble Or Boom?

Are we in an AI bubble? Parallels are being drawn to the dot.com boom/bust of 1999-2000.

In the dot.com bust, many high-tech companies valuations soared up 10X, then deflated. The peak P/E ratio for the Nasdaq Composite was 200! Remember Webvan? It went public November 1999 with an $8 billion valuation, then filed for bankruptcy 19 months later. It was much speculation without profits or growth.


Fig. 1: ChatGPT image – AI Bubble vs. AI Boom

Today’s valuations are up significantly, but so are revenues, earnings and growth prospects. However, market enthusiasm is leaking into companies with risks like Coreweave.

The Nasdaq Composite P/E ratio at the end of September is 33. Amazon, Apple, Google, Microsoft, Meta and TSMC are in the 27-39 range. Nvidia is rich at 52, but they are hugely profitable and the #1 winner from AI, and will continue to be for some time. AMD’s 97 P/E ratio is a bet that they can grab significant share in GPUs in the next few years. As of this writing, it has jumped to 140 P/E based on their OpenAI deal. Only Tesla, at 265 P/E, is in the dot.com bubble range, losing on EVs, lagging Waymo badly on autonomous driving, and betting big on humanoid robots with little evidence.

GenAI revenues are accelerating rapidly. TSMC, in its recent earnings call, said demand for data center AI grew significantly since last quarter and they are struggling to keep up. GenAI has real and compelling value for numerous enterprise applications. In April, McKinsey projected $5 trillion will be invested in AI data centers through 2030. Much of this is coming from the world’s largest and most profitable companies. But much is also coming from companies with highly leveraged balance sheets. GenAI will get much bigger. How much bigger remains to be seen. And then, who wins and who loses?

Adoption of GenAI/LLMs is rapid

ChatGPT took only 2 months to reach 100 million users at launch. TikTok took 1 year; WhatsApp 3 years; Facebook 4.5 years; Twitter 5 years; LinkedIn 8 years, and NetFlix 10 years (Source: BOND Capital and company filings).

OpenAI’s revenues have soared to $4.3 billion for 1H 2024. Projections for 2030 are as high as $200 billion, which equals half of Apple’s current revenues. Open AI has been announcing deals at a steady clip and is locking up lots of GPU capacity, building their own data centers, their own AI accelerator, and announcing numerous new products to add value, grow TAM, and increase stickiness. OpenAI expects to have positive cash flow in 2029 at $125 billion in revenues. It is betting on rapid growth to pay for its numerous large commitments that apparently total about $1 trillion. In fact, Sam Altman appears to be making commitments to lock in resources that he believes will be scarce, but necessary to achieve the growth.


Fig. 2: OpenAI Revenue. Source: The Information

Anthropic announced their Series F last month and disclosed that at the start of this year their annualized run-rate revenue was $1 billion/year and had grown to $5 billion/year by August.

Google, meanwhile, processed 10 trillion tokens in April 2024. By June 25, that number had climbed to 980 trillion. That’s a doubling every 2 months! Google doesn’t disclose GenAI revenue separately.

Morgan Stanley last month computed Microsoft’s AI revenue (implied from CapEx and Margins) to be at least $12 billion in FY2025, and projected it to be at least $85 billion in FY29.

Part of this rapid adoption is a combination of increasingly better models (one-shot to reasoning to agentic) combined with increasingly lower costs for AI (driven by a combination of better GPUs, bigger scale up networks, and improvements in LLM architectures). As the chart below shows, AI inference pricing has come down more than 100X in just 2 years:


Fig. 3: AI inference price for customers per 1 million tokens, 11/22 to 12/24. Source: Stanford HAI

This incredibly fast price drop enables the shift from one-shot models to reasoning and agentic AI, which require 1 to 2 orders of magnitude more compute. Advances in GPUs and the size of scale-up networks will further enable the ability to improve models and cut costs simultaneously.

There is a learning curve to using these models. Prompt engineering is a valued skill today. To get value out of these models takes effort and persistence for top results. Those who learn to deploy AI effectively will gain share in their industries versus their competitors who lag on AI.

GenAI adoption is real and historically rapid – can it continue?

Consumer adoption of GenAI is rapid

Menlo Ventures in June said 60% of American adults used AI in the last 6 months, and 20% rely on it every day. A $12 billion consumer AI market has been built in 2.5 years. That revenue is coming from a small percentage of the users because most are using free versions. The free versions get people hooked. Over time more of them will pay the $20/month to get the more advanced models, faster response time, and voice mode (which is great for practicing languages and interaction).

The firm found that young people use AI the most, but it also found that 10% of daily users are seniors. The survey shows ChatGPT is the top AI assistant, followed by Gemini, Meta, Alexa, and Siri. Consumers have practically no switching cost and find all of the AI assistants to be similar.

Most expected Google’s search business would be gutted by Chatbots. They have lost some search share, but their integration of Gemini with search is masterful (and is resulting in huge traffic drops at sites that thrived on links instead of AI-generated answers).

OpenAI is now introducing “stickier” products like Sora 2 video generation and being able to view and buy products in the ChatGPT app. Look for more of these differentiated offerings.

The consumer market will grow, but the enterprise market is ultimately much bigger.

Enterprise use cases of GenAI are huge and disruptive

Enterprise will be the biggest market for GenAI. This is the focus for Anthropic and most of the hyperscalers.

Today’s GenAI has significant value. The best models can match the performance of top humans on things like international math Olympiads. DeepMind’s AlphaFold can predict the structure of nearly all known proteins, which was considered super hard just five years ago. And AI does especially well at boosting the productivity of coders — so much so that most companies don’t need to hire more.

Some of the early use cases for GenAI in the enterprise (ARR = annual recurring revenue run rate):

  • Cursor AI code editor: ARR from $0 in early 2023 to $300M/year April this year (Claude, ChatGPT, and others find software coding to be a big use case too).
  • Lovable no-code building of web sites and apps: started late 2024, $50M/year ARR 5/25.
  • Abridge Healthcare Clinical Conversations: $50M 10/24 doubling to $100M 3/25.
  • Harvey Legal Workflows: $10M ARR in 12/23 to $70M ARR in 3/25.
  • Decagon Customer Service AI Support: $1M in 2023 to $10M in 2024.
  • Alphasense Financial Research & Analysis: $200M ARR in 2022 to $400M ARR 2024.

There are many more applications in development. Look at Microsoft Discovery Agentic AI for R&D. I recently saw a video at a conference that showed how numerous specialized agents cooperated to solve a complex thermal mechanical analysis in a couple hours. It would have taken months for humans using EDA tools.

Some major companies are finding AI can enables them to grow without hiring. Examples:

  • Walmart’s CEO recently stated AI is reshaping every job, and that he is freezing hiring because AI enables them to continue to grow revenue without adding people.
  • Salesforce’s CEO announced earlier this year that they would stop hiring software engineers, citing the efficiency gains from AI.
  • Amazon’s CEO said mid-year that they will need fewer people doing some of the jobs being done today due to GenAI.

I recently met two founders of a company providing tools to speed marking tests and papers at high school. They expect they can reach a $1 million/year run rate before they need to hire a third employee.

Walmart has 2.1 million employees with an estimated annual payroll of $79 billion. They are growing revenues about 4% annually to $695 billion/year. Using AI to boost productivity of their current team and avoid hiring might save about $5 billion in annual payroll in a few years.

Current GenAI has large value, and we are still low on the learning curve for how to deploy it. Moreover, GenAI will keep getting better as models and hardware advance.

How many GenAI winners will there be?

Training frontier models costs billions of dollars, so the number of players will be limited to those with hyperscale resources. In November 2024 Anthropic’s CEO foresaw data centers growing in size from $1 billion in 2024 to $10 billion in 2026, then to $100 billion in 2027-28. The larger the data center required to train frontier models, the fewer players that can afford to enter the game.

Embedding models into cloud ecosystems (Azure, AWS, Google Cloud) or social/consumer platforms (Meta, Apple) creates lock-in advantages. Also, some firms like Google can let their models access lots of customer files that others cannot, giving them an edge and “stickiness.” As the chart below shows, ChatGPT with 800 million users is super impressive, but the Hyperscalers and Apple have even more users that are very invested in their ecosystems.


Fig. 4: Incumbents’ advantage. Source: Bond Capital

Another source of competitive advantage for the large GenAI companies is developing their own XPU AI Accelerators. Amazon AWS and Google started early and have large deployments.

GPUs are able to run any workload. But an optimized XPU only has to run the workloads for that company, and it can be focused just on inference, which is rapidly becoming the dominant share. So the GenAI companies will be able to run their workloads on their XPUs at lower power and cost for a given throughput and responsiveness. For training, and for their cloud customer workloads that run on CUDA, they will use GPUs.

An indicator of XPU market share is Global CoWoS (advanced packaging) capacity demand by customer. Morgan Stanley 7/17/2025 estimated that total demand for 2026 is almost 900K wafers, with Nvidia taking ~2/3 of the total; the other major players being Broadcom at 10% and Alchip/AWS and AMD both at 5%. So about 15%-20% of CoWoS unit share is for proprietary XPUs.

Bond Capital estimates that sales of Google TPUs and Amazon GPUs are growing fast to significant levels. See the charts below:


Fig. 5: Google TPU sales. Source: Bond Capital


Fig. 6: Amazon Trainium sales. Source: Bond Capital

The market for AI Accelerators is growing to $500 Million+/year, according to AMD’s Lisa Su. Nvidia and AMD are already top 10 customers for TSMC. GenAI companies like AWS, Google, and OpenAI are likely to become top 10 customers for TSMC (directly or indirectly). As their XPU revenues grow, they may further integrate key semiconductor suppliers through acquisition.

We are likely to see multiple winners in GenAI because OpenAI, Anthropic, and the major Hyperscalers are vertically integrating across models, ecosystems, data centers, and proprietary XPUs. Customers will find an ecosystem that meets their needs and be reluctant to change.


Fig. 7: The largest GenAI players are vertically integrating.

There are other possible players on the fringes with LLMs and/or proprietary XPUs: xAI, Cerebras, and major Chinese players. They could surprise the market and join the top-tier club. But it will be challenging to keep up with the largest GenAI players.

The best strategy for the smaller ones is regional or industry specialization. Cerebras, for example, benefits from its wafer-scale processor having more memory on a single chip, enabling it to run GenAI models with better throughputs. It has opened 7 data centers across the US and just opened it’s first in Europe.

The most vulnerable companies are Oracle and Coreweave. They have a currently valuable asset (scarce GPUs), but don’t have their own LLM, ecosystem, or broad customer base. Their customers may desert them when GPUs become plentiful.

The most likely structure by 2030 is:

  • Tier 1 Frontier Leaders: 3-6 of Anthropic, AWS, Google, Meta, Microsoft, OpenA, and maybe one Chinese giant driving cutting-edge research.
  • Tier 2 Clouds & Integrators: Players like Apple, Oracle, Salesforce, and those in the list above that don’t make Tier 2 building value-added vertical LLMs.
  • Tier 3 Regional: Mistral in Europe, Alibaba/Baidu/Bytedance/Tencent in China.
  • Tier 4 Niche Specialists: Domain-optimized models for industries.

What could go wrong?

Risk No. 1: Demand soars, but lags data center capacity
The major Hyperscalers are funding their expansion out of their immense cash flow. A shortfall would be painful but manageable.

OpenAI and Anthropic are funding their expansion out of unprecedented fundraising in the private markets. Their deep-pocketed investors would be hurt if down rounds were required, but not fatally.

The companies with the most risk in this scenario are firms like Coreweave and Oracle – they don’t have their own workloads to run and they don’t have a huge base of customers. Hyperscalers/OpenAI/Anthropic would prioritize using their own data centers and drop Coreweave and Oracle as soon as their contracts would allow. Oracle and Coreweave have significant debt funding. But it’s a small percentage of the total industry CapEx.

Risk No. 2: One or two players dominate instead of 5+
If GenAI is a commodity and ecosystems are less critical than the best model, then one or two players could dominate by spending more money than others can keep up with and take most of the market. In this scenario, it is possible they need so much compute that they contract to use the capacity of the laggards rather than those data centers being empty.

It seems there is enough opportunity for differentiation, specialization, and bundling with ecosystems that this is not the most likely scenario.

Risk No. 3: Hardware doesn’t deliver expected improvements as much/soon as expected
Nvidia has an extremely impressive roadmap requiring a handful of new complex chips every year. If this slipped, the ability to run larger models at lower costs would suffer and could delay a ramp for economics. OR Hyperscalers could shift to AMD and/or their own accelerators to make up.

Risk No. 4: Fab & packaging capacity lags demand
TSMC essentially makes all of the AI accelerators and related complex logic chips in the AI data center. Recent reports indicate TSMC is raising wafer prices as much as 50% for the 2nm node, which presumably will be used for the newest AI chips coming to market. TSMC has been an extremely effective and reliable partner and is most likely using this money to fund fab capacity to continue to meet the growth needs of the major AI players — or at least what they are willing to fund. TSMC is building about 10 new fabs and 2 new advanced packaging facilities, in Taiwan and Arizona, with a total spend of around $40 billion.

Samsung and Intel are the only other options for advanced nodes in advanced packaging. So far, they have not won (publicly at least) any major AI accelerator commitments other than Tenstorrent. If fab capacity is squeezed, they may become an option, but their current capacity is relatively small compared to TSMC for the advanced nodes.

Risk No. 5: Power lags demand
At the Morgan Stanley Technology Conference in March, Sam Altman said that he expects the limiter of AI expansion to be power. The Information recently reported that an internal memo at OpenAI outlined Sam’s thinking that they should have 250 GW of data centers in 2033. That’s a fourth of the total power generated in the U.S. today, and the same as India’s. (It also would cost more than $10 trillion; the entire industry’s CapEx is only $3 trillion to $4 trillion through 2030).

Data centers have responsiveness requirements, so they can’t all be in one spot serving the world. Instead they are spread across the US and the world. They need to be located close enough to serve their region with the necessary responsiveness but also where they can get the power needed.

The demand for power is huge. Meta is building a 5GW data center in Louisiana, which is about the same amount of energy used for all of Los Angeles. A 1 GW data center will have an annual power bill in the ballpark of $50 million/year.

For decades the US electric grid didn’t grow. Increased consumer demand was offset by increased energy efficiency. That is changing fast with AI. Data centers used 2% of US electricity for many years until 2020. In 2024, that increased to 4.5%, and it will grow much larger. Consumer electricity rates in areas serving large data centers are increasing significantly.

A 1 GW solar installation would require 12.5 million solar panels, covering nearly 5,000 football fields, which equals 8 square miles (source: New York Times 9/29/25). Only remote locations are feasible for this. And the cost of batteries is huge, and the availability is limited. Wind is more problematic.

Nuclear plants are being restarted under contracts with hyperscalers. OKLO (market cap $17 billion) has non-binding LOIs for over 1 GW worth of power from Modular Nuclear Reactors that can be deployed at the data center location.

Clean energy isn’t enough. With current plans, most data centers will run on natural gas — if they can get turbines. They’re sold out through 2030.

Getting enough energy online fast enough will get harder if the ramp accelerates. The good news regarding a bubble is the data center companies aren’t likely to build and buy chips without having the power commitments first.

The AI power crunch is an opportunity for AI semiconductor solutions that deliver more AI per watt.

Conclusion

The winners will have the best models, the best responsiveness, and the lowest costs. It might be 1 or 2 winners, but more likely there will be 5+ because of differentiation and ecosystem bundling.

Semiconductor companies should focus on delivering what the winners need to win. Being conservative could mean missing the next Industrial Revolution.

The post AI Bubble Or Boom? appeared first on Semiconductor Engineering.

wpChatIcon
wpChatIcon