The US tech sector - particularly the mega caps - are the only sector in the world you “HAVE” to have a view on. When I look at US tech, I see the following (please feel free to abuse me in the comments). Most of the tech space is exactly as you would imagine it to be - volatile and competitive. However big tech - Microsoft, Meta, Google, Apple and Amazon have managed to carve out profitable areas (I will come to Nvidia later). Microsoft has managed to become the “go to” corporate software name. A huge installed base, and large switching costs, allows it to generate significant cash flow. The shift to cloud has helped to reduce sales on original software, and expansion into other areas. While growth is good, it has had to invest to grow.
Google has a dominant position in search. This has allowed it to generate huge growth in cashflow with far less capex than Microsoft until recently.
Meta another advertising giant has been able to growth cashflow with little resort to Capex.
Apple barely bothers with Capex at all.
And then there is Amazon, who basically spends everything its makes to grow.
Each has a monopoly in various area - Google in search, Meta in social media, Apple in handsets, Microsoft in corporate software, and Amazon in online retail. By far the best business is search (social media is good too - but as Tiktok shows, can be either disrupted, or banned). Its such a good market, Google paid Apple USD 20bn a year just not to launch their own search engine. The arrival of AI - Open AI and now DeepSeek - has thrown open the search market. Rather than just asking Google - asking your selected AI program is the default. For Google this is an existential threat, and hence Google is investing USD 75bn. Microsoft, Meta and Amazon are not far behind leading to a USD265bn capex in total on AI. For me, the amount of spending is not just to defend existing market positions, but also to discourage start ups from entering the space entirely. If you set AI bar at USD 50bn to compete - then most of your competition disappears. One way to do that is to bid up the value of Nvidia GPUs - and Nvidia has surged due to this tactic.
In simple terms, Google is trying to defend its position in search, and everyone is else is trying to get in. With limited number of Nvidia chips available, this makes perfect sense. This is why DeepSeek is such a problem - as it seems to imply this tactic is broken. That a competitive AI model can be built much more cheaply, and without cutting edge chips. The question is whether the big tech companies decide that spending on AI makes sense or not. Assuming AI will be a advertising led model (DeepSeek makes that more likely), we can estimate profitability. We can see at how much Meta makes per subscriber. In the most recent quarter than was USD 14.25 on its 3.35bn subscribers or about USD 50 per sub a year.
With USD 265bn Capex from the big 4, and lets assume that they get USD 50 per sub - would mean 5.3bn users (i.e the whole internet connected world) just to get the capex back in sales - and that is ignoring running costs - which are substantial. ChatGPT obviously thinks it can charge directly - it wants USD 20 a month from average punters, and USD 200 a month from corporates. A month ago, I was prepared to pay this - no longer.
If AI is going to be free - and DeepSeek makes this more likely - then something has to give. My guess is that should a rally in commodity prices, or a sell off in bonds tighten financial conditions - AI budgets will be the first to get cut. For the moment, financial conditions remain benign.
The risk to this view is that DeepSeek is a fraud, or the AI market changes again - but given the Capex numbers we are talking, and the availability of DeepSeek as a free option in the AI space - a cut in AI capex seem inevitable to me.
Sustained sell off in bonds are always a function of rallying commodity prices. Therefore the short MAG7 thesis, even with this perspective, still requires higher oil prices to trigger the retrenchment in capex. Once again, commodity prices are the only real Kryptonite for MAG7.
Oil won't go higher (in the USA) unless natural gas also goes higher. Oil is VERY overpriced compared with natgas on an energy-equivalent basis right now.
Today, Natural Gas Prices in Europe are trading at $100 WTI Equivalent. EU companies are substituting oil products for NG right now IF (Big IF) they have the ability to switch fuels and all companies are contemplating making the investment to be able to use the cheapest BTU available in the future. It will take time and US investment in LNG export infrastructure, but if you ultimately are a believer in BTU Price Convergence, oil will ultimately trade with Global NG on a BTU equivalent basis. Lastly, the AI Datacenter build out will drive electricity prices higher and NG will go along for the ride as it is currently the best fuel source (reliable with no interuptability 24/7/365 unlike Alt Fuels) for datacenters until there is a breakthrough in small-scale nuclear module reactors. All that said, this is the long view.
Yes - NG in Europe is very expensive at the moment. But what I would also say is the food prices are also kryptonite to markets - and having spiked in 2021/2 - have flatlined since, but not collapsed. A surge in food inflation also seems possible to me.
"Lastly, the AI Datacenter build out will drive electricity prices higher"
I disagree here.
Future data centers won't be grid-tied. They will have their own combined cycle natgas turbines. Gas goes in and data comes out with no effect on local electric rates.
I also see data centers being built in colder climates to reduce cooling costs. A significant % off data center costs are just keeping the server cool.
Both are being done - 1) Datacenters being powered through the electricity grid with contacts already signed and in pace and messing with residental electricity prices even though they say they isolate it out - BS! and 2) The major oil companies (XOM, CVX, etc - Best alternative IMHO) bypassing and setting up gas-powered plants right on the location of the datacenter. You are correct, ultimately the second route will predominate, but there are many Datacenter plans already contingent on tying into the electricity grid as the second and more logical and efficient method came to the pary late. Why were the Major oil companies late to the game - they wanted Iron Clad contracts are they see this whole AI BS as I do (next paragraph)
To be 100% honest I think AI is a horrible investment and we will never achieve an ROIC > WACC. We will never reach AGI and I think we have will have hundreds of billions of stranded investments. Just look at OpenAI's Financial Statements (via The Information, FT, and WSJ - if it can be trusted.) One of these days the AI companies will try to raise more cash as they continue to burn through past cash raises and they will not be able to round out the book. Capex to NVDA and the rest of the picks and shovels will quickly dry up and we see a combination of the 00/01 Tech Bubble and 07/08 GFC (With Housing being replaced by vast datacenters and other peripheral AI assets as their loans secured by NVDA Chips and the like will soon be pennies on the dollar.) We will find the use cases are quite minimal and we WAY OVERSPENT! How is that for a bear story! This is the worst broad investment hype cycle (and magnitude) I have seen in my almost 40 years playing this stupid, foolish game!
I love a good rant! More likely it will be like all new technogloies - overhyped in the short term, understimated in the long term. Internet and shale are two good examples of both
"if you ultimately are a believer in BTU Price Convergence"
I am. NG should trade for about 6x the price of oil on an energy equivalent basis, and 10x oil taking into account the extra costs of handling a gas.
Henry hub $4.15/mmbtu and WTI $73.33 as of this morning... BIG difference to arbitrage away.
Only thing I would say is that this is happening on a global level - just that as the US is an exporter, its domestic gas price has to reflect transport costs - so is lower than world price
True, but there are other aspects pushing down natgas in the US...
Due to recent(ish) regulations, natgas produced as a byproduct of oil extraction can no longer be flared off, it has to be piped out. Because there is finite pipeline capacity at the oil wells, the excess gas (and natgas liquids, NGLs) limits how much oil can be extracted.
In other words, we have limited supply of oil but excessive supply of natgas. This is noteworthy because this is purely a regulatory restriction, not physical or geological.
NB: my interpretation of exploration surveys is that we have an effectively infinite supply of natgas for the next 50+ years.
perplexity has installed a version of deepseek after scrubbing it of the chinese censorship. i doubt it's fake.
while I don't doubt that deep seek performs as advertised, I think the ultimate question is did they actually build it for $5.6mm, or is that number the deception and they used the latest chips, sourced through resales from Singapore, and spent $100 mm. After all, the Chinese are known to have lied about things in the past, and this would be a great way to attack the US equity market without much effort
Chinese tech across almost all fields has come in much cheaper. Have they benefited from seeing other people develop tech first - undoubtedly - but so has every other country since the first industrial revolution.
The key IMO is not how much it costed to train but how much it costs per token for inference. If indeed it is 1/20 of OpenAi and & this is indeed quite bearish for LLMs providers
My understanding is that they were quite clever with the software to keep costs down. And the idea of country having to innovate around sanctions is not that unsual. South Africa developed technology to turn coal to petrol during its sanction period.
they never said they built it for that amount. they said they trained it for that amount.
but if it still cost billions to build, and they don't allocate the cost while the Western firms do, is that the right comparison? it could certainly be a loss leader, but it also changes the context of the conversation.
I'm no expert, just want to consider that not everything China says is reality, but rather their narrative
If distillation reduces costs - it does imply even new models in the west should be much cheaper - and with Meta going open source - then surely DeepSeek proves AI should be much cheaper than advertised
i have read that deepseek was trained as a distillation of openai. i am not qualified to know whether this technique allows for progress in functionality, or just efficiency in development. that is, are distilled models limited to the functional level of the foundational models on which they're trained, and is progress in function dependent on foundational models alone. if so, there's a problem that foundational models are expensive but can never earn back their cost.
I have no view on Nvidia but the trend is towards on device AI. Eg meta's rayban and versions of their Orion later this year. More on-phone AI. (That said, I do have a long Meta short Apple pair trade on)
I like the idea - and the Rayban tech has improved a lot. I thought SnapChat was also doing something with glasses
New versions of the Meta wearable likely have a superior silicon battery from Enovix (https://www.globenewswire.com/news-release/2025/01/07/3005436/0/en/Enovix-Secures-Landmark-Purchase-Order-for-Silicon-Batteries.html)
One of the board members of $envx comes from $meta. Zuckerberg last month criticized apple (https://finance.yahoo.com/news/mark-zuckerberg-slams-apple-innovation-002230094.html)
https://finance.yahoo.com/news/mark-zuckerberg-slams-apple-innovation-002230094.html
https://research.gavekal.com/article/tariffs-and-the-platform-company-model/
The question I have is what part of the Mag7 valuation is coming from AI hype? I could see CapEx stay elevated then come down as an equilibrium of "good enough" and "maintenance CapEx" is determined. For example, MSFT already is offering co-pilot for free and their lock-in for corporate software means that they just need to be "good enough". https://www.cummulative.io/p/why-did-microsoft-give-github-copilot
One observation I have is that this view has been discussed a lot elsewhere. Shouldn't it already be priced into the market?
It is an interesting question. But I see Mag 7 as representing costs to corporate America - so as long as sales are growing, they can grow revenue. Conversely, once capex starts to get cut - revenue expectations fall, just as Capex is on the rise. Deepseek is just the beginning - most US software seems overpriced to me - its just that switching costs are hassle - but a recession would probably seen churn rise
great post but the French do not call it AI they call it IA=Intelligence artificielle, always special our French friends :-)
I prefer the French pronunciation of "Russell" to the English version as well!
what is the difference between FARP and Sub ? why USD 50bn per sub ?
Some heroic assumptions that DeekSeep moves AI from a subscription model to an add based model The facebook numbers are quarterly, so I annualised to get USD 50 a sub
There are a number of other threats to nvdia out there. There is a good analysis here https://youtubetranscriptoptimizer.com/blog/05_the_short_case_for_nvda?mc_cid=3205a8b633
So even if capex remains strong generally, the market will push the hardware back to the commodity model and a lot of capital value will be lost. The simplest way will be if the companies have to shorten their depreciation periods.