A couple of comments here are circling the power question, so one angle from the infrastructure side.
In the State of the AI Economy deck, slide 36 carries the cost of consumed electricity: the $594m energy line, which is reasonable on its own. What it does not separately identify is the cost and lead time to make a GW of IT load actually available at the site. On-site power and cooling sit in the facility line, but utility-side interconnection, transmission, substations, queue timing, and fixed-capacity commitments appear only inside a $33m "land + utility" line, if at all.
That matters because large-load power is not purely usage-based. Demand charges, minimum-billing demand, and take-or-pay commitments keep much of the bill fixed even when utilization runs below plan. And slides 32-34 frame "headroom" as revenue after depreciation, not after OpEx, so energy is in the slide 36 unit cost but outside the "paying back" claim.
Slide 16's data center examples show both paths. Rainier in New Carlisle takes the grid route: a ~2.3 GW build phased over years, with a $150m+ transmission and 345kV substation buildout. xAI's Colossus took the other route, self-supplying ~1.2 GW of gas off-grid because the interconnection queue was too slow; per the SpaceX IPO filings it now runs ~1 GW of IT compute across three buildings, well past the chart's early-2025, single-building 300 MW marker.
Forward-looking, the swing on cost per token is less the marginal electricity price and more when, how, and under what fixed commitments a site gets powered.
Hi Corey! Yes, I think your lead-time point is very relevant: buying compute and building the DC doesn't help if it isn't connected and you haven't planned for behind-the-meter power generation. And delays put even more pressure on expended CapEx as it's spent without generating revenue.
On slide 36 we account for the substation CapEx depreciation charge in the land + utility line, and it constitutes $20m of the total $33m annual cost (we had to balance brevity/clutter on this so couldn't be as comprehensive in describing as we wanted to!). But this doesn't cover interconnection queue position, transmission upgrades beyond the substation, or the fixed-capacity commitments because they are site- and utility-specific (hard to represent in a stylised steady-state model).
And you're exactly correct on the headroom point: this is just to cover the pure depreciation charge. The remaining 19%/32% of revenues (slide 34) currently has to cover all DC OpEx as well as all the costs of generating value-add (from the foundation model & app layers) on top.
Thanks for coming back on this in detail. Good to know the $20m is the substation depreciation, and I get the brevity-versus-clutter tradeoff makes sense for this type of model. Agreed that interconnection position, transmission beyond the substation, and fixed-capacity terms are genuinely site-specific and hard to represent in a steady-state model. That's the part I find most interesting going forward.
Seems to me that the risk to the Frontier Models has been dramatically increased by the US Govt blocking access to Fable/Mythos 5.
The Chinese OpenWeight models are catching fast and priced at 10% of the Frontier and as majority of work is good enough on Opus or DeepSeek - the desire for most organisations to leverage Frontier (except Defence and high frequency FS) could decline rapidly.
The Energy issue is huge and IMHO is a battle only the Chinese can win. The buildout of US Data Center capacity is coming under increasing pressure from all over the country and when there is talk of 9GW Datacenters covering areas larger than NYC and using more power than the rest of the state, does it not become obvious that the exponential growth will run over a cliff in 2027, if not sooner.
Thanks! First study of this kind I've seen; looking forward to future reports and trend analysis. Some of my encounters with AI enhanced interfaces are very iffy. Is there away in the future to separate out user learning curve improvements?
Thank you so much. Can you help explain how we should think about the impact of power costs embedded in this analysis, including on a forward looking basis? I need to take a closer read, but is it already embedded inside the token costs or are these incorporated elsewhere?
Thanks! Best place for power costs is in slide 36, where we look at the unit cost of tokens: energy costs are 2/3 of datacentre OpEx, but CapEx costs are the most significant driver (89%) of the token cost floor
Hi all from the Exponential View team, would anyone have initial thoughts on Kimi K3 model, and importantly linking to the broader implications from your report? Put differently how should we think conceptually about the demand side of the equation?
Of course there is still a lot we dont know but any initial thoughts would be much appreciated.
Firstly: What excellent work! I haven't seen anyone beyond a few very sell-side investment banks try to put this together, and what I have seen hasn't drawn on data so well. Congratulations to the EV team. Keep it up!
I have a couple of follow up points / questions that I'd be delighted if someone could address:
1. Chinese models: A large share of US AI traffic (close to a majority?) now seems to flow to Chinese models — reportedly ~60% of US startups use them, and Microsoft is talking up DeepSeek alongside its own. Your slide 60 shows OpenRouter token share shifting to open-weight, and Scenario 1 (slide 62) assumes labs lose the frontier licensing premium. I'd be interested in your thoughts on how the "loss" of revenue to heavily subsidised Chinese models (IP theft, cheap energy, other indirect subsidies) affects the high-spending US/UK frontier labs (OAI, Anthropic, GDM) and the US AI economy more broadly. How do the frontier labs become profitable against that competition, a US ban on Chinese models? Compliance and privacy moats? If the US government acts to protect AI revenues (and effectively the sole source of economic growth in the US right now), how does this affect things?
2. Chip overhang? Given the slow build of data centers vs announced plans, what happens to previous-generation, unused chips (e.g. Blackwells) when Vera Rubin comes online? Do they still get deployed as sites eventually finish? Does anyone have a handle on how many are sitting waiting — I've seen estimates from "a few hundred thousand" to "2 million". That seems like a dangerous balance sheet item for someone, and it bears directly on the 6-year useful-life assumption doing so much work in slides 37–38. If they sit idle for a year, does their revenue generating useful life reduce to 5 years? Does this hold if the next generation gets installed first? Had a few conversations with Epoch people myself on the depreciation.
3. Interest cost: Unhedged today (8 July) summarised BoE's version of the BIS's "AI might be due a correction" piece. Beyond valuations, the striking part is the projected collapse in hyperscaler free cash flow (caveat Meta, which seems to have decided it has more compute than it needs — not sure they continue much beyond the planned build). We've all seen signs of it already of course. Given the switch to bond issuance (~$110bn in 2025; H1 2026 has already exceeded that), I'd love clarity on slide 36: the note says the $7.9bn/GW cost is annualized including cost of capital — but at what rate, and does it reflect the marginal dollar increasingly being debt-funded (your slide 31)? Even ~4% on the cumulative issuance is $10bn+/yr of interest and compounding. I believe they are paying something like 6%?
4. The OAI IPO paradox: This one is really stumping me. The leaked (FT-verified) OAI accounts show ~$8bn of actual 2025 cash burn against $13bn revenue — the scary $38.5bn headline was mostly a one-off non-cash charge (although if you are paying staff in stock so heavily, after IPO does that remain the case?). But the financing structure around OpenAI's previous private round makes the IPO load-bearing: SoftBank's $40bn bridge loan comes due March 2027, and $35bn of Amazon's cheque is gated on a listing (or verified AGI!!!) before end-2028 per SEC filings. Meanwhile banks reportedly wouldn't lend against OpenAI paper without a SoftBank parent guarantee — the $852bn private mark isn't yet bankable, at least if you are SoftBank, despite your golden goose. I know, OpenAI isn't "the AI economy," but it's a critical hub for both demand (analysts put ~$300bn of Oracle's $638bn RPO with OpenAI — relevant to your slide 12 backlog chart) and financing. So my question here for EV: if the IPO slips to 2027 and the funding chain tightens, how much of your demand growth and CapEx trajectory (slides 28–31) is exposed to this single counterparty? Is there a concentration-risk view in the model? Also interested in your takes on why OAI seems to be delaying its IPO?
The usable unit of value slide is the best one - question remains on what will be left for the rest of us.
Act as System Integrators? Harness Designers?
Several questions to be discovered and a lot to be developed towards the downstream.
cheers!
rogério
much to explore
A couple of comments here are circling the power question, so one angle from the infrastructure side.
In the State of the AI Economy deck, slide 36 carries the cost of consumed electricity: the $594m energy line, which is reasonable on its own. What it does not separately identify is the cost and lead time to make a GW of IT load actually available at the site. On-site power and cooling sit in the facility line, but utility-side interconnection, transmission, substations, queue timing, and fixed-capacity commitments appear only inside a $33m "land + utility" line, if at all.
That matters because large-load power is not purely usage-based. Demand charges, minimum-billing demand, and take-or-pay commitments keep much of the bill fixed even when utilization runs below plan. And slides 32-34 frame "headroom" as revenue after depreciation, not after OpEx, so energy is in the slide 36 unit cost but outside the "paying back" claim.
Slide 16's data center examples show both paths. Rainier in New Carlisle takes the grid route: a ~2.3 GW build phased over years, with a $150m+ transmission and 345kV substation buildout. xAI's Colossus took the other route, self-supplying ~1.2 GW of gas off-grid because the interconnection queue was too slow; per the SpaceX IPO filings it now runs ~1 GW of IT compute across three buildings, well past the chart's early-2025, single-building 300 MW marker.
Forward-looking, the swing on cost per token is less the marginal electricity price and more when, how, and under what fixed commitments a site gets powered.
Hi Corey! Yes, I think your lead-time point is very relevant: buying compute and building the DC doesn't help if it isn't connected and you haven't planned for behind-the-meter power generation. And delays put even more pressure on expended CapEx as it's spent without generating revenue.
On slide 36 we account for the substation CapEx depreciation charge in the land + utility line, and it constitutes $20m of the total $33m annual cost (we had to balance brevity/clutter on this so couldn't be as comprehensive in describing as we wanted to!). But this doesn't cover interconnection queue position, transmission upgrades beyond the substation, or the fixed-capacity commitments because they are site- and utility-specific (hard to represent in a stylised steady-state model).
And you're exactly correct on the headroom point: this is just to cover the pure depreciation charge. The remaining 19%/32% of revenues (slide 34) currently has to cover all DC OpEx as well as all the costs of generating value-add (from the foundation model & app layers) on top.
Thanks for coming back on this in detail. Good to know the $20m is the substation depreciation, and I get the brevity-versus-clutter tradeoff makes sense for this type of model. Agreed that interconnection position, transmission beyond the substation, and fixed-capacity terms are genuinely site-specific and hard to represent in a steady-state model. That's the part I find most interesting going forward.
This is an amazing analysis.
Seems to me that the risk to the Frontier Models has been dramatically increased by the US Govt blocking access to Fable/Mythos 5.
The Chinese OpenWeight models are catching fast and priced at 10% of the Frontier and as majority of work is good enough on Opus or DeepSeek - the desire for most organisations to leverage Frontier (except Defence and high frequency FS) could decline rapidly.
The Energy issue is huge and IMHO is a battle only the Chinese can win. The buildout of US Data Center capacity is coming under increasing pressure from all over the country and when there is talk of 9GW Datacenters covering areas larger than NYC and using more power than the rest of the state, does it not become obvious that the exponential growth will run over a cliff in 2027, if not sooner.
I think this could very well be a risk that will need to be managed.
Thanks! First study of this kind I've seen; looking forward to future reports and trend analysis. Some of my encounters with AI enhanced interfaces are very iffy. Is there away in the future to separate out user learning curve improvements?
Thank you so much. Can you help explain how we should think about the impact of power costs embedded in this analysis, including on a forward looking basis? I need to take a closer read, but is it already embedded inside the token costs or are these incorporated elsewhere?
Thanks! Best place for power costs is in slide 36, where we look at the unit cost of tokens: energy costs are 2/3 of datacentre OpEx, but CapEx costs are the most significant driver (89%) of the token cost floor
Thank you! Super interesting
Hi all from the Exponential View team, would anyone have initial thoughts on Kimi K3 model, and importantly linking to the broader implications from your report? Put differently how should we think conceptually about the demand side of the equation?
Of course there is still a lot we dont know but any initial thoughts would be much appreciated.
Firstly: What excellent work! I haven't seen anyone beyond a few very sell-side investment banks try to put this together, and what I have seen hasn't drawn on data so well. Congratulations to the EV team. Keep it up!
I have a couple of follow up points / questions that I'd be delighted if someone could address:
1. Chinese models: A large share of US AI traffic (close to a majority?) now seems to flow to Chinese models — reportedly ~60% of US startups use them, and Microsoft is talking up DeepSeek alongside its own. Your slide 60 shows OpenRouter token share shifting to open-weight, and Scenario 1 (slide 62) assumes labs lose the frontier licensing premium. I'd be interested in your thoughts on how the "loss" of revenue to heavily subsidised Chinese models (IP theft, cheap energy, other indirect subsidies) affects the high-spending US/UK frontier labs (OAI, Anthropic, GDM) and the US AI economy more broadly. How do the frontier labs become profitable against that competition, a US ban on Chinese models? Compliance and privacy moats? If the US government acts to protect AI revenues (and effectively the sole source of economic growth in the US right now), how does this affect things?
2. Chip overhang? Given the slow build of data centers vs announced plans, what happens to previous-generation, unused chips (e.g. Blackwells) when Vera Rubin comes online? Do they still get deployed as sites eventually finish? Does anyone have a handle on how many are sitting waiting — I've seen estimates from "a few hundred thousand" to "2 million". That seems like a dangerous balance sheet item for someone, and it bears directly on the 6-year useful-life assumption doing so much work in slides 37–38. If they sit idle for a year, does their revenue generating useful life reduce to 5 years? Does this hold if the next generation gets installed first? Had a few conversations with Epoch people myself on the depreciation.
3. Interest cost: Unhedged today (8 July) summarised BoE's version of the BIS's "AI might be due a correction" piece. Beyond valuations, the striking part is the projected collapse in hyperscaler free cash flow (caveat Meta, which seems to have decided it has more compute than it needs — not sure they continue much beyond the planned build). We've all seen signs of it already of course. Given the switch to bond issuance (~$110bn in 2025; H1 2026 has already exceeded that), I'd love clarity on slide 36: the note says the $7.9bn/GW cost is annualized including cost of capital — but at what rate, and does it reflect the marginal dollar increasingly being debt-funded (your slide 31)? Even ~4% on the cumulative issuance is $10bn+/yr of interest and compounding. I believe they are paying something like 6%?
4. The OAI IPO paradox: This one is really stumping me. The leaked (FT-verified) OAI accounts show ~$8bn of actual 2025 cash burn against $13bn revenue — the scary $38.5bn headline was mostly a one-off non-cash charge (although if you are paying staff in stock so heavily, after IPO does that remain the case?). But the financing structure around OpenAI's previous private round makes the IPO load-bearing: SoftBank's $40bn bridge loan comes due March 2027, and $35bn of Amazon's cheque is gated on a listing (or verified AGI!!!) before end-2028 per SEC filings. Meanwhile banks reportedly wouldn't lend against OpenAI paper without a SoftBank parent guarantee — the $852bn private mark isn't yet bankable, at least if you are SoftBank, despite your golden goose. I know, OpenAI isn't "the AI economy," but it's a critical hub for both demand (analysts put ~$300bn of Oracle's $638bn RPO with OpenAI — relevant to your slide 12 backlog chart) and financing. So my question here for EV: if the IPO slips to 2027 and the funding chain tightens, how much of your demand growth and CapEx trajectory (slides 28–31) is exposed to this single counterparty? Is there a concentration-risk view in the model? Also interested in your takes on why OAI seems to be delaying its IPO?