Very nice read! Would note that CFO capex initiatives commonly span across multiple years and so the cumulative spend should be considered when completing the ROI math; additionally, worth noting that most growth capex spend will require some level of maintenance capex as well. Agree with the notion of cumulative revenues to be used in the ROI math, though, as you pointed out, AI-specific rev will be hard to pinpoint due to knock-on effect of ancillary revenue streams and potential cost savings.
Thanks, ‘Digestion’ seems a fair way to state the near future. The only analogue really is the telecom boom and bust of 2000. The capabilities were used but not on a schedule that made investors money. At present many of these AI capabilities are in low margin usages and like many things on the web, users will expect them to be nearly free or won’t use them. At present AI revenues are much less than Americans spend on pet food. I looked that up on Google answers… because it was free. These capabilities will find uses in time but that will take partners and specificities. The everything flavor of LLM seems a bet against long odds.
1. all capex is to support GPU rollout, GPUs are not long term assets
2. your spend numbers don’t come close to sama’s ask
I get there is nuance, but the bull case for AI is definitely close to the top.
and I get the hype. but the shift towards RL and other techniques means the raw token output is not of sufficient quality and will create downwards pressure on margins.
It’s too much work to do my own business case calculations to show you are totally off base. But some qualitative observations make the point.
First, you omit the fact that these capex investments are not for a single year. They have been happening for years now, and all indications are that they will happen for years to come. So, it’s wrong to compare aggregate cash flow revenue to a single year of investment.
Second, the only revenue that matters for these business cases is what comes in to the companies making the capital investments, and which are attributable directly to the investments. Benefits to users, satellite enterprises, or society as a whole don’t count. Neither do “soft” projections, like future cost avoidance or possible increases in ad revenue. Veterans of the business case game know such attributions are pure fudges. They never really come to pass.
If you apply these principles to the calculations, the investments make zero sense.
The premise is “$560B capex for AI” with no time dimension. Are you claiming there has been $0.5T/yr capex each of the last several years?
Obviously there has been spending, but that looks like a cumulative number to me.
Also “claims of efficiencies are BS” ok sure in the “unspecified synergies from merger” style, but this seems similar to electrifying a steam manufacturing plant or connecting branch offices to HQ with a leased line. Sure it seems amorphous, but five years on try to take it away and you’ll see the value.
An excellent piece . Informative and thoughtful . Question is what will Capex be in 2026 and 2027 ? Any forecast ? Will the MaG 7 be okay with paltry short term returns in exchange for seeing up the market now ?
Thanks, Feisal. I think 2026 is more “digestion” than collapse: capex stays big but growth slows, with dollars remixing toward training pods, memory-rich nodes, and custom silicon. 2027 depends on utilization. if fleets are running hot, spend can re-accelerate; if not, it flattens further. And yes, the Mag 7 seem fine with thin short-term returns if it buys them long-term moats.
Very nice read! Would note that CFO capex initiatives commonly span across multiple years and so the cumulative spend should be considered when completing the ROI math; additionally, worth noting that most growth capex spend will require some level of maintenance capex as well. Agree with the notion of cumulative revenues to be used in the ROI math, though, as you pointed out, AI-specific rev will be hard to pinpoint due to knock-on effect of ancillary revenue streams and potential cost savings.
Thanks, all good points!
Thanks, ‘Digestion’ seems a fair way to state the near future. The only analogue really is the telecom boom and bust of 2000. The capabilities were used but not on a schedule that made investors money. At present many of these AI capabilities are in low margin usages and like many things on the web, users will expect them to be nearly free or won’t use them. At present AI revenues are much less than Americans spend on pet food. I looked that up on Google answers… because it was free. These capabilities will find uses in time but that will take partners and specificities. The everything flavor of LLM seems a bet against long odds.
I don’t know.
1. all capex is to support GPU rollout, GPUs are not long term assets
2. your spend numbers don’t come close to sama’s ask
I get there is nuance, but the bull case for AI is definitely close to the top.
and I get the hype. but the shift towards RL and other techniques means the raw token output is not of sufficient quality and will create downwards pressure on margins.
It’s too much work to do my own business case calculations to show you are totally off base. But some qualitative observations make the point.
First, you omit the fact that these capex investments are not for a single year. They have been happening for years now, and all indications are that they will happen for years to come. So, it’s wrong to compare aggregate cash flow revenue to a single year of investment.
Second, the only revenue that matters for these business cases is what comes in to the companies making the capital investments, and which are attributable directly to the investments. Benefits to users, satellite enterprises, or society as a whole don’t count. Neither do “soft” projections, like future cost avoidance or possible increases in ad revenue. Veterans of the business case game know such attributions are pure fudges. They never really come to pass.
If you apply these principles to the calculations, the investments make zero sense.
You have to do a bit more than that.
The premise is “$560B capex for AI” with no time dimension. Are you claiming there has been $0.5T/yr capex each of the last several years?
Obviously there has been spending, but that looks like a cumulative number to me.
Also “claims of efficiencies are BS” ok sure in the “unspecified synergies from merger” style, but this seems similar to electrifying a steam manufacturing plant or connecting branch offices to HQ with a leased line. Sure it seems amorphous, but five years on try to take it away and you’ll see the value.
An excellent piece . Informative and thoughtful . Question is what will Capex be in 2026 and 2027 ? Any forecast ? Will the MaG 7 be okay with paltry short term returns in exchange for seeing up the market now ?
Thanks, Feisal. I think 2026 is more “digestion” than collapse: capex stays big but growth slows, with dollars remixing toward training pods, memory-rich nodes, and custom silicon. 2027 depends on utilization. if fleets are running hot, spend can re-accelerate; if not, it flattens further. And yes, the Mag 7 seem fine with thin short-term returns if it buys them long-term moats.