Bear market prices for a historic growth period
Investors and the broader public seem pessimistic/dismissive about AI adoption in the years ahead and may get surprised about the reality as before.
The stock market is discounting the growth of AI stocks, they are priced as if their sales and revenue growth will stop next year, maybe even decline. The demand signals show the opposite, HBM is sold out for the year and getting into Q1, memory makers are even getting guaranteed buying contracts for years in advance with minimum pricing specified.
Nvidia’s forward price/earnings ratio is as low as it has been in the 2022 bear market (~20), when their sales were declining and the entire market was undervalued due to sentiment. Currently, they are more than doubling revenue and maintaining margins. Growth is expected to slow down next year, but we are far from declines, and estimates are usually low. Some other names like memory makers are more richly valued but many hyperscalers are similarly discounted. Normally, tech stocks like these trade at 30-70 P/E, sometimes even higher.
The bear thesis
The market is discounting growth and is down for multiple reasons:
Hyperscaler capex is at unprecedented levels, it is the first case in a long time that they need to use debt/share offerings to get more capital as their free cash flow is turning negative. This is contrary to years of buying back shares and accumulating cash on the balance sheet.
Frontier model lab revenue growth is slowing down. This was bound to happen, previous growth rates were unsustainable (>1000%/year does not last for long, ever)
Average token price (“token index”) started to come down as enterprises started to correct the overspending. The “tokenmaxing” era of random unlimited experimentation is over. It made sense when costs were insignificant, it doesn’t now.
Nvidia is getting more and more creative with vendor financing, even guaranteeing loans and GPU usage for neoclouds. This does not raise their risk significantly, they have $80B in cash and $13B debt.
News of Chinese chip fab advancements raise concerns of competition. Their DUV technology is years behind EUV and can produce DRAM and logic chips for edge devices. They cannot produce HBM or advanced chips comparable to TSMC or the Korean memory makers. In a few years, they might be able to catch up in technology, but ramping the production will probably take longer. Their domestic market will take most of their production for the foreseeable future which is already off limits for the rest of the chipmakers due to sanctions, so even if they succeed in catching up (big if), they are not likely to meaningfully compete in international markets until the end of the decade.
Short term noise
These are also scaring the markets:
Mass margin calls after retail FOMO (fear of missing out), mostly concentrated in Korean memory makers by locals. Millions of accounts were liquidated, causing a temporary but harsh selloff. Trading was halted on the Korean stock exchange 7 times so far this year by circuit breakers due to similar selloffs. Price drives sentiment, retail traders often sell after, extending drops.
Kimi K3 release, a near-frontier open weight model from China. This seems to be a mirror of the Deepseek R1 release from last year. Having cheaper tools will not decrease revenue when market penetration is very low and quickly growing. It could if the market was saturated and demand was inelastic. Neither is true.
Fears of FED rate hikes usually cause selloffs the tech sector. Inflation and oil price spikes have some investors concerned. Even if the FED hiked, it wouldn’t stop the wave of investment. The productivity increases coming from the technology are overwhelming, 0.25% or 0.5% higher interest rates wouldn’t make a dent in the momentum.
Growth signals
The reasons growth is not likely to stop:
There are lots of catalysts for growth and they are a lot stronger:
Agentic model use consumes an order of magnitude more tokens than interactive chatbot use. Most companies are barely scratching the surface of this use case. The TAM (total addressable market) is enormous and the market penetration is minimal. It’s not just coding, personal assistant type bots are useful for everyone. Long-running research tasks can be done in every domain. I have been using it for statistical analysis on trading strategies. Lots of traders are not coders. You can also do online research for sentiment analysis or brand analysis based on comments etc. This wave has already started and is in full swing, already showing significant results. Coding agent adoption is easy, the rest is progressing more slowly.Video generation seems to be on the cusp of being able to generate entire movies and is still improving. Costs are also dropping, there are smaller and smaller models with the same quality, but it still uses a lot more tokens than text. Replacing Hollywood movies is not the TAM, although even those studios are using more and more special effects, generating more and more scenes and characters, but when it becomes orders of magnitudes cheaper to create entire movies, we’ll have a lot more of them. Indie creators are going to flourish. We’ll need more GPUs.
Robotics is starting to ramp up. It will take a couple of years to have a significant impact but it is already starting. It will need a lot of simulation for synthetic data generation, which also uses GPUs. But then all that training needs to happen and even more chips are needed for inference on the edge. Nvidia has been talking about this for a while and developing specific chips for edge devices. Self-driving is a similar use case and requires similar chips for local inference. They also need a lot of memory to run the models and store the context. Robotics training is very hard (compute intensive) due to the high variability and length of tasks to be performed and also due to the massive amount of input video and sensor data in addition to the high optionality of actuator movements.
Summary
There are multiple secular trends that underpin growth and none of the scary signals the market is reacting to are measuring up in impact. They are at most, bumps in the road, some of them are just misunderstandings, like the token index (average token price) dropping.
We are likely to have shortages in semiconductors in general, more so in memory and energy for the foreseeable future, a few years at least. The signs that could show a slowdown, like datacenter rental prices or single GPU hourly prices have been increasing, even for older chips like the H100. This means useful life and return on investment for chips is better than expected.
This could all drop together if demand fell, memory and GPU prices would fall, smaller companies would be at risk, but the large ones like Nvidia and the hyperscalers have no bankruptcy risk, only the risk of falling profits. Many of them are already priced for that, and there are no signs of demand falling. This is a good setup for upside surprises.