It has been a deep reddish black November.
Datacenter buildouts got pushbacks post mid-terms. Jensen’s affairs, his subsequent departure, and possibility of return, and the rumor of a US intelligence op, have made tabloids out of most technology publications. Then came the Hunterbrook short report on AMD about China shipments. The AI trade got obliterated. Everybody is depressed and much poorer than a few months ago.
So where are we really in AI now?
TomorrowX
Everyone was searching for the answer at TomorrowX held in Austin TX, an arguably unfortunately timed event. It was smack dab in the middle of the reddest week in recent memory. Still, kudos to Gavin Baker and team for being forward-thinking and putting this together. The idea behind the event is amazing: put independent voices on Substack/X in the same room as investors and tech companies, and help everyone drive the narrative of AI together.
Sentiment on the floor was doom and gloom. Jensen departing from Nvidia was viscerally affecting some attendees. Others were reeling from their holdings dropping significantly, or even going negative. Many investors rotated out of AI semis and infra, into other sectors. Some investors claimed that they de-grossed in July but I only have to take them at their word.
With all the fear, it is difficult not to wonder if this is a buying opportunity. Should we try to catch this falling knife? Will we see a pullback in the earnings of companies supplying into the AI infra space? Are capex investments going to slow down? SemiAnalysis recently projected that CapEx will go from ~$1.5T to about $900B by 2030 — a roughly 40% downward revision. Sudden revisions are hard to take at face value considering that they might revise it upward just as quickly.
Attendees at TomorrowX hoped that the legendary investor Stanley Druckenmiller, CEO of Duquesne Family Office, would have some answers. You should listen to the whole talk. The following quote about what Druckenmiller would do in this environment stood out (emphasis mine).
So, my choice here is to do what I’ve always done when I don’t have clarity, and I’ve been doing this, oh my God, for fifty years I’ve been doing this. I can’t remember a period where I had less clarity on where we’re gonna be in twelve months, unlike last January. So as you made fun of me on one of these Zoom calls the other day, you said, “Stan has a button now. When I call, it just says, ‘I’m not playing.’
It’s clear we are in crazy times where uncertainty and volatility reigns supreme. One can naturally expect the outlook in 2027 to be conservative, which is arguably good for the AI infrastructure trade. Expectations had gotten out of line with reality, and the current market could really be oversold as well.
Early Innings
For what its worth, we are still in the early stages of what is a revolutionary technology whose deployment, adoption, and eventual benefits takes its time. Meanwhile market/macro/mistress events often distract from the marvel of what we have today. So let us remind ourselves.
AI has made math essentially a solved problem. The Riemann Hypothesis was an unsolved problem for >150 years, and all of a sudden two solutions appear on the same day (1, 2 - sourced from Exponential View). Less than two months ago, OpenAI reported a solution to the Navier-Stokes existence and smoothness problem. How many more solutions will we see in 2027 to Millennium Prize problems? Much to the chagrin of mathematicians, they have been reduced to mere verifiers of AI generated solutions. FUNDA wrote an excellent column on this last month that I recommend reading.
AI is actively being applied to the sciences too, and Druckenmiller at the same conference said that he expects cancer to be a solved problem in the next 10 years. Whether that will happen or not, we now have a new tool — like PCR was to Genomics — that we can use to accelerate discovery of a cure. Maybe we can find low-cost cures to rampant diseases in poor countries and uplift all of humanity. Someone asked me at TomorrowX how I am so chipper given the state of markets; it is because I believe we have an endless road of possibilities ahead of us.
In my opinion, our early innings has only seen chatbots and computer-based knowledge work. We still have not unlocked the true capabilities of what this technology is capable of. Railroads unlocked big box stores, the latter of which provided outsized market returns for decades compared to the underlying infrastructure itself. I do not believe we have identified the true beneficiary of AI yet.
To be clear, I do not consider myself AGI-pilled. There does not be a manifestation of god-like intelligence for this technology to be useful — just like we do not need the grand unified theory of physics to build useful machines. I am a believer in smaller pockets of useful intelligence tuned to specific outcomes: a biology model, a space-research model, a chip-design model, etc. These do not have to be small or open-source models. They can be multi-trillion parameter models built by large corporations with domain expertise to accelerate progress. That is more bullish for compute and the overall economy than the AGI scenario.
AI Infrastructure Trade
The days of bottleneck hunting are over, and is not a frame investors can use anymore. Order backlogs into 2028-29 mean nothing if supply does not come online to convert backlogs to revenue. Anything can happen over longer time frames (or even in weeks, as we have seen lately), so putting too much emphasis on the demand for a particular company’s products does not automatically make them a good investment.
If the current situation does provide a rare buying opportunity, then where should investors look? Well, there are a few options.
For a large part, TSMC has been conservative with CapEx and commentators like Ben Thompson have pointed out that this could be much to their own detriment if demand explodes. In fact, TSMC has been happy to let competitors have some business if they are not able to meet the demand, especially as related to advanced packaging. That fiscal conservativeness might just save us from entering a chip supply glut going forward. TSMC is one of those companies that fit the profile of a company with a durable moat, sustained demand, and solid execution. Intel, unfortunately, is none of these things today.
Making chips requires wafer fab equipment (WFE), and with Elon promising to disrupt chip manufacturing with TeraFab, this sector holds quite some promise going forward. Elon said at TomorrowX that they expect their first test chip in 2027, and volume production by end of 2028 for phase one. Rumors are that TeraFab is targeting 1 million logic wafers and 700K memory wafers, per month, by 2032. A 100 million square foot facility as promised by TeraFab will require a lot of WFE. We are yet to see this promise materialize of course, but we are only looking for long term directional trends and WFE is directionally aligned with need for more compute.
Energy is a finite resource and if we have exponential compute needs, then power-efficiency is everything. Our human brain consumes about 20W, and with 8B people on the planet, our current biological compute demand costs about 160GW. Current projections for AI datacenter power demand are more than twice that by 2030. Something needs to give here because simply deploying a million GPUs in a datacenter with each one consuming 2-3KW of power is just brute force and likely not sustainable. I have argued that constraints on land and power availability makes efficiency the highest priority of the datacenter buildout going forward.
Luckily, efficiency improvements come from the entire AI stack, and that implies opportunities to innovate at every level.
GPU/XPU: Calculations required for AI must consume lower power by moving less data around, adopting lower power technologies, or finding innovative architectures.
Memory: Going to vertically stacked memory on logic reduces data movement distance, and consequently energy use to move bits. In my conversation with the Micron CTO, Scott DeBoer, at TomorrowX, he explained that this is the highest priority for the company as there are many customers asking for it.
Networking: Minimize energy cost of moving bits around by adopting near/co-packaged optics. Slow and wide approaches with micro-emitters are promising, and optics for scale-in would be a massive performance unlock too.
Algorithms and Software: Finding ways to quantize more effectively, reduce computations, or orchestration via software to improve energy utilization still holds promise. Lots of work is still going on here.
There are just some examples and I am sure there are many more. The main point is that our need for compute is still growing, and power demands are growing with it. We will eventually adapt towards getting more compute out of every watt deployed, and I believe a lot of investing opportunities lie in that directional trend.
How to Think About This?
It is easy to talk positively about AI in a bull market, and be right most of the time. Virtually any AI-related stock in the last few years has made money for investors. Today’s conditions require a combination of conviction, a solid understanding of technology, and an open mind to find investing opportunities going forward. The AI semis trade will continue to provide returns in the future, but not at the frenetic pace that we have seen in the past few years. I view that as a good thing because we will transition to an investing environment where rational thinking translates to good returns. Hype/fear-driven cycles are much harder to invest in.
While AI semis will continue to be attractive past this temporary phase of contraction, the next sequence of multibaggers will likely arise from outside it. To identify those opportunities, it is important to be on the lookout for AI’s RealBeneficiary™.
To quote Druckenmiller again (emphasis mine):
In terms of equities, we are looking for AI use-cases, as opposed to infrastructure. Infrastructure may be a great tactical trade because there have been really washed out positions. I don’t think they are going to be the engine they were to provide the scoop they did for us the last four years, but I think use-cases could.
Ironically, the most interesting use-case is a lab. And that’s because I don’t personally they are going to be in the token business. My own belief is that they are not going to be releasing models, and if you have the number one intelligence in the world, you can start selling tasks instead of tokens.
Tasks, not tokens is definitely a reframe I will be watching out for. In the near term, it is best to set out to work identifying what semi companies have taken a greater beating than they should have. Those could be useful investing opportunities in semis.




