Point-by-Point Review: Dylan Patel — Deep dive on the 3 big bottlenecks to scaling AI compute

This is a point-by-point review of Dylan Patel — Deep dive on the 3 big bottlenecks to scaling AI compute against our analysis. Dylan Patel is CEO of SemiAnalysis, a semiconductor and AI infrastructure research firm. The podcast is hosted by Dwarkesh Patel. Claims are evaluated against our analysis of orbital AI datacenter economic competitiveness vs terrestrial alternatives.

Summary

Overall assessment: The Patel source is overwhelmingly consistent with our analysis. Of 32 identified points, 16 directly align with positions and evidence already cited in our pages. The source is already well-integrated into our analysis — we cite 6 specific evidence items from it across the gpu-useful-life, orbital-operational-lifetime, and other pages.

The 4 novel supporting evidence points strengthen our conclusions through new framing or data not currently cited: the reliability disadvantage of space optical links vs. terrestrial transceivers (Point 14), data center modularization accelerating terrestrial deployment (Point 18), the 20x inference performance gap between GPU generations for MoE models (Point 22), and failure management as a competitive differentiator impossible to replicate in orbit (Point 23).

The 2 "merits investigation" points both relate to chip supply constraints as a separate and independent argument against orbital compute (Points 7 and 25). These do not challenge our TCO conclusions but identify a structural argument we have not fully developed: in a chip-constrained world, the question of where to deploy GPUs is dominated by deployment speed and cost, both of which favor terrestrial. This could warrant a brief section in cost-parity-timeline noting that the chip supply ceiling makes the orbital energy advantage moot when energy is not the binding constraint on total compute capacity.

Merits investigation

Point 7: Chip manufacturing is the binding constraint, not power

Patel argues that by 2028-2029, the binding constraint shifts from power/data centers to chip manufacturing. He traces this to ASML EUV tool production being limited to ~100 tools/year by 2030, with each gigawatt of compute requiring about 3.5 EUV tools. The total chip manufacturing capacity by end-of-decade is ~200 GW.

Merits investigation. Our analysis treats GPU hardware cost and availability as exogenous — the same chips are available for both orbital and terrestrial deployment. But Patel's argument has a deeper implication: if chips are the binding constraint (not power or land), then the orbital value proposition (unlimited free energy, no land permitting) addresses the wrong bottleneck entirely. Our cost-parity-timeline discussion of "What the Ratio Range Implies for the Business Case" mentions capacity deployment speed and energy sovereignty as potential justifications for an orbital premium, but does not address the chip constraint. If chip supply caps total compute deployment at ~200 GW/year regardless of where it is deployed, the orbital argument weakens further because:

• The urgency to deploy at speed is capped by chip availability, not facility availability
• The "option value" of faster orbital deployment is eliminated if chips arrive on the same schedule regardless
• The cost premium for orbital is borne on a chip-constrained resource that could generate higher returns terrestrially

This does not change our TCO conclusion (orbital is more expensive in all scenarios), but it adds a powerful additional argument against the orbital business case that our analysis currently lacks. The affected page would be cost-parity-timeline, specifically the "What the Ratio Range Implies for the Business Case" section. Research needed: validate Patel's ~200 GW/year end-of-decade chip production ceiling and assess whether chip supply constrains total compute deployment more tightly than power supply.

Point 25: ASML EUV tools as the ultimate constraint — only ~100/year by 2030

Patel provides detailed analysis of ASML's supply chain (Cymer sources, Zeiss optics, reticle stages) and argues that even with aggressive expansion, ASML can only produce ~100 EUV tools/year by 2030. At 3.5 tools per GW, this yields ~200 GW total chip capacity by decade's end.

Merits investigation. While the absolute GW ceiling does not change the orbital-terrestrial cost comparison directly (both compete for the same chips), it has an important second-order implication for our analysis: if total annual chip production is capped at ~200 GW, then the total addressable market for orbital compute is bounded by the fraction of those chips that orbital operators can capture. Our cost-parity-timeline analysis currently discusses cost competitiveness and workload feasibility but does not address chip allocation as a constraint. In a chip-constrained world, the question is not just "is orbital cheaper?" but "can orbital operators even get chip allocation when terrestrial operators have higher willingness to pay due to faster time-to-revenue?" This warrants at least a qualitative note in the cost-parity-timeline discussion. Research needed: assess whether chip allocation preferences (faster terrestrial deployment -> faster revenue -> higher willingness to pay for allocation) create a structural disadvantage for orbital operators beyond the cost premium.

Addressed — we reach a different conclusion

Point 17: GPU value increases over time due to more capable models ↗

Patel argues "an H100 is worth more today than it was three years ago" because improvements in model capability (GPT-5.4 vs GPT-4) mean each GPU produces more economic value. This counteracts the normal depreciation narrative.

Addressed — we reach a different conclusion. While Patel's observation about GPU value per model improvement is correct for market pricing in a capacity-constrained environment, our analysis treats GPU useful life as an economic depreciation concept (5 years central), not a value-per-token concept. For the orbital vs. terrestrial TCO comparison, what matters is the cost of the hardware ($/kW_IT) and how long it operates, not the revenue it generates. Both orbital and terrestrial deployments benefit equally from model improvements (the same GPU running a better model earns more regardless of location), so this effect is neutral in the orbital-terrestrial comparison. We do note in gpu-useful-life that chip manufacturing constraints could extend useful life — Patel's "value increases" argument is the demand-side version of this supply-side observation.

Novel supporting evidence

Point 14: Space networking is more expensive and unreliable than terrestrial pluggable transceivers ↗

Patel observes: "In a cluster, 15-20% of the cost is networking. All of a sudden, you're using space lasers instead of simple lasers that are manufactured in volumes of millions with pluggable transceivers. And those things are very unreliable as well, more unreliable than the GPUs by the way. Across the life of a cluster, you have to unplug and clean them all the time."

Novel supporting evidence. Our analysis addresses the bandwidth gap between ISLs and NVLink but does not explicitly address the reliability differential between terrestrial pluggable transceivers and space optical links. The observation that terrestrial optical transceivers require frequent physical maintenance (unplugging, cleaning) — which is impossible in orbit — adds a reliability dimension to the networking constraint beyond pure bandwidth. This supports our broader conclusion that orbital compute faces maintenance challenges that terrestrial deployment does not, reinforcing the effective lifetime penalty discussed in orbital-operational-lifetime.

Point 18: Modularization of data centers reduces construction bottleneck ↗

Patel describes a trend toward factory-integrated data center modules: entire rows of servers, networking, cooling, and power shipped as pre-assembled skids from Asian factories. He argues this will "drastically reduce the number of people working in data centers, so our capability to build them will be much larger."

Novel supporting evidence. Our terrestrial-energy-supply-constraints page discusses supply-side dynamics for terrestrial power but does not specifically address data center construction modularization as a factor that accelerates terrestrial deployment speed. This trend strengthens the terrestrial case: if modular, factory-built data center components can be deployed in weeks rather than months, the speed advantage that orbital proponents claim (deploy satellites faster than building data centers) is further eroded. This is relevant to the "Capacity deployment speed" discussion in cost-parity-timeline.

Point 22: The performance gap between Hopper and Blackwell is much larger than FLOPS difference — ~20x for inference ↗

Patel explains that for inference at 100 tokens/sec on DeepSeek and Kimi K2.5, "the performance difference between Hopper and Blackwell is on the order of 20x." This is not just the 2-3x FLOPS improvement but compounds NVLink bandwidth, memory bandwidth, and on-chip communication advantages.

Novel supporting evidence. Our analysis notes the 3-4x generational performance improvement but does not cite the 20x inference-specific difference between Hopper and Blackwell for MoE models. This supports our conclusion that orbital hardware becomes obsolete rapidly — a satellite launched with Hopper-era hardware delivers inference at 1/20th the throughput of a terrestrial Blackwell deployment for the same model. This compounds the effective lifetime penalty in orbital-operational-lifetime and further undermines the orbital value proposition: orbital hardware is not just 3-4x behind each generation but potentially 10-20x behind for the specific workloads (MoE inference) that matter most.

Point 23: Cloud provider differentiation comes from managing GPU failures — impossible to replicate in space ↗

Patel explains: "We have ClusterMAX, which rates all the neoclouds. We test over 40 cloud companies... Outside of software, what differentiates these clouds the most is their ability to deploy and manage failure." He notes neoclouds take up to 6 months to deploy GPUs on Earth, and asks: "Where does space get in there?"

Novel supporting evidence. Our analysis addresses hardware failure rates (15% Blackwell RMA, ~6% annual GPU failure rate) and the inability to perform physical repairs in orbit, but does not frame failure management as a competitive differentiator that orbital operators fundamentally cannot match. Patel's observation adds a service-quality dimension: terrestrial operators compete on how well they handle the inevitable failures (rapid RMA, hot-swap, sparing), while orbital operators have no equivalent capability. This reinforces the broader argument in in-orbit-servicing-feasibility that software-managed graceful degradation is the only near-term failure management strategy for orbital compute.

Consistent

Point 1: H100 TCO and GPU deployment economics ↗

Patel states that "an H100 costs $1.40/hour to deploy at volume across five years" and that signing a deal at $2/hour yields roughly 35% gross margin. He notes H100 pricing fell from ~$2/hour (2024) to ~$1/hour (2026) as Blackwell deployed at volume.

Consistent. These figures are already cited in our analysis. The GPU useful life page (gpu-useful-life) references patel-2024-ai-bottlenecks.1 for the $1.40/hour TCO figure at 5-year depreciation and patel-2024-ai-bottlenecks.3 for the price decline trajectory. These data points anchor our GPU cost modeling and support the 5-year central useful life estimate.

Point 2: GPU depreciation cycle — 5 years is the standard assumption ↗

Patel uses 5 years as the standard GPU useful life assumption: "If your argument is that a GPU has a useful life of five years..." He dismisses Michael Burry's argument for 3-year depreciation as overly bearish, while noting that in a chip-constrained world, depreciation could extend beyond 5 years because older GPUs retain value.

Consistent. Our gpu-useful-life page uses a central estimate of 5 years, matching Patel's assumption. We cite patel-2024-ai-bottlenecks.4 and patel-2024-ai-bottlenecks.5 for these positions. The countervailing force Patel identifies — chip manufacturing constraints could extend GPU economic life — is explicitly noted in our analysis under "Trend direction: slight shortening."

Point 3: NVIDIA 2-year cadence with 3-4x performance improvement per generation ↗

Patel describes the NVIDIA generational cadence: "every two years NVIDIA is tripling or quadrupling the performance while only 2X-ing or 50% increasing the price." He maps out Hopper -> Blackwell -> Rubin with each generation delivering roughly 3x performance.

Consistent. Our gpu-useful-life page references patel-2024-ai-bottlenecks.2 for this claim and uses it to derive the rapid depreciation trajectory: "After two generations (4 years), older hardware is 9-16x less efficient per watt." This is a key input to the economic obsolescence dimension of orbital operational lifetime.

Point 4: Testing and deployment delay for orbital GPUs consumes 10% of useful life ↗

Patel argues: "If your argument is that a GPU has a useful life of five years, and this takes six additional months, that is 10% of your cluster's useful life." He emphasizes that in a capacity-constrained world, compute is theoretically most valuable in the first months you have it.

Consistent. We cite this argument directly at patel-2024-ai-bottlenecks.1 in the orbital-operational-lifetime page and the gpu-useful-life page. The latter states: "Any time spent on ground testing, launch, and orbital commissioning (estimated 3-6 months by Dylan Patel) reduces the productive window by 5-10%." The argument that compute is most valuable early, when capacity constraints are tightest, strengthens the case against orbital deployment where the lead time before first revenue is longer.

Point 5: 15% of deployed Blackwells require RMA ↗

Patel states: "Even today, around 15% of Blackwells that get deployed have to be RMA'd. You have to take them out."

Consistent. Our gpu-useful-life page notes: "Terrestrial data centers can swap individual GPUs (15% RMA rate for Blackwell)." This is one of the key terrestrial advantages over orbital: component-level repair is routine on Earth but impossible in orbit without in-orbit servicing (which we assess as speculative and unproven for LEO component-level repair in in-orbit-servicing-feasibility).

Point 6: Space GPUs are not happening this decade ↗

Patel argues explicitly: "Space data centers effectively are not limited by their energy advantage. They are limited by the same contended resource. We can only make two hundred gigawatts of chips a year by the end of the decade. What are we going to do to get that capacity? It doesn't matter if it's on land or in space." He concludes: "I don't see how you could test them all on Earth, deconstruct them, and ship them to space without it taking significantly longer than just leaving them in the facility where you tested them."

Consistent. Our cost-parity-timeline analysis reaches the same conclusion: orbital compute does not reach cost parity with terrestrial in any scenario or time horizon. The chip supply constraint argument Patel makes is captured in our analysis as a deployment speed concern: "The optimistic scenario's 2030 ratio may be economically correct but practically irrelevant — multi-GW commercial orbital compute deployment by 2030 is implausible." Our reasoning is cost-driven (persistent premium) while Patel adds a chip-supply-constraint argument that reinforces the same conclusion through a different mechanism.

Point 8: Energy is only 10-15% of GPU cluster TCO; doubling power cost adds only ~$0.10/hour to Hopper cost ↗

Patel states: "Even if power costs double on Earth, it's still a fraction of the total cost of the GPU." He calculates: "The Hopper that was $1.40 is now $1.50 in cost. I don't care, because the models are improving so fast that the marginal utility of them is worth way more than that ten-cent increase in energy."

Consistent. Our cost-parity-timeline analysis reaches the same structural conclusion: "GPU cost is common to both sides and dominates both TCOs" (74% of terrestrial central TCO, 65-70% of orbital amortized capex). The energy cost analysis in terrestrial-energy-cost shows central energy at ~$694/kW_IT/year — roughly 6-7% of total terrestrial TCO. Eliminating this through orbital solar saves ~$580-$720/kW_IT/year, which our analysis explicitly finds "insufficient to close the gap." Patel's $0.10/hour increase framing ($876/year per GPU) is consistent with our finding that energy cost is a second-order factor.

Point 9: Power scaling is not the bottleneck — diverse BTM generation can deliver hundreds of GW ↗

Patel describes an extensive ecosystem of power sources: combined-cycle gas turbines (three major manufacturers), aeroderivative turbines (airplane engine conversions), reciprocating engines (Cummins-class diesel/gas), ship engines, Bloom Energy fuel cells, solar plus battery, and various novel entrants. He states: "Any of these individually will do tens of gigawatts, and as a whole, they will do hundreds of gigawatts." He tracks 16 different power generation equipment vendors with "hundreds of gigawatts of orders."

Consistent. Our terrestrial-energy-supply-constraints side page documents this in detail, including specific manufacturers and cost figures. We estimate total feasible annual U.S. supply addition at ~30-55 GW/yr by 2028-2030. The terrestrial-energy-cost page tracks the cost trajectory of these diverse sources. Patel's characterization aligns with our central scenario: supply constraints cause temporary cost elevation in 2028-2032, then ease as diverse BTM solutions scale. The key data point — 16 vendors tracked by SemiAnalysis with hundreds of GW in orders — provides additional backing for our position that terrestrial power scarcity is not a structural constraint.

Point 10: BTM gas can be as expensive as $3,500/kW and it barely moves TCO ↗

Patel argues: "It can be as high as $3,500 per kilowatt. It could be twice as much as the cost of combined-cycle, and the total cost of the GPU on a TCO basis has only gone up a few cents per hour."

Consistent. Our terrestrial-power-asset-capex and terrestrial-energy-cost pages model BTM gas costs ranging from $800-$1,800/kW_gen for simple-cycle to higher for aeroderivative/reciprocating. Patel's $3,500/kW upper bound exceeds our modeled range but confirms our structural finding: even expensive BTM power has negligible impact on total GPU TCO because energy is a small fraction. This further undermines the orbital argument that "free" solar energy in space offsets other cost premiums.

Point 11: Permitting is a challenge but solvable — America has 50 states ↗

Patel pushes back on the permitting/land argument for space GPUs: "Land-wise, America is big. Data centers don't actually take up that much space... Dealing with permitting and red tape in middle-of-nowhere Texas, Wyoming, or New Mexico is probably a hell of a lot easier than sending stuff into space." He notes the Trump administration made permitting easier, and that Texas in particular allows skipping red tape.

Consistent. Our cost-parity-timeline acknowledges the orbital deployment speed argument but notes the terrestrial counterargument: "behind-the-meter (BTM) generation is the market's actual response to grid constraints. 56 GW of BTM generation is already planned for US data centers, deployable in 6-12 months at costs below orbital." Patel's point about geographic diversity of US siting options reinforces that land/permitting is not a structural constraint favoring orbital deployment.

Point 12: Networking requirements make inter-satellite inference impractical for frontier models ↗

Patel argues that inference for leading models requires hundreds of GPUs communicating at terabytes per second within a scale-up domain: "DeepSeek's production deployment... they were running on 160 GPUs." He notes three levels of bandwidth: on-chip (tens of TB/s), within-rack via NVLink (TB/s), and between racks (hundreds of Gbps). On Starlink inter-satellite links at 100 Gbps, he says: "But that's per GPU, not per rack. So multiply that by 72. Also, that was Hopper. When you go to Blackwell and Rubin, that 2x's and 2x's again."

Consistent. Our inference-networking-requirements page contains a detailed analysis of exactly this bandwidth gap. We quantify it as an "~18x per-endpoint gap" between a single 800 Gbps ISL and NVLink 5 (1.8 TB/s per GPU), and note that matching the NVL72's 130 TB/s aggregate all-to-all bandwidth would require "~1,300 equivalent ISL links, which is topologically infeasible for a satellite cluster." Our tiered feasibility framework (Tier 1: single-satellite, Tier 2: small-cluster PP, Tier 3: wide-EP infeasible) captures Patel's argument precisely.

Point 13: Sparse MoE models drive need for wide expert parallelism across many GPUs ↗

Patel explains that as models become more sparse (MoE), "you want to ping just a couple of experts per GPU. If leading models today have hundreds, if not a thousand, of experts, then you'd want to run this across hundreds or thousands of chips." He emphasizes this compounds the inter-satellite networking problem.

Consistent. Our inference-networking-requirements page documents this trend in detail, citing nvidia-wide-ep-nvl72.1 showing EP=32 achieves 1.8x more throughput than EP=8, and nvidia-moe-frontier-models.2 showing over 60% of frontier open-source models use MoE. Our analysis concludes: "MoE architectures dominate frontier models, and MoE inference benefits enormously from wide EP requiring large NVLink domains." This trend toward wider expert parallelism makes the orbital networking constraint more binding over time, not less.

Point 15: Higher watts per mm^2 is more difficult in space — cooling in space is harder, not easier ↗

Patel argues: "If manufacturing is the constraint, right now it's roughly one watt per square millimeter for AI chips. One easy way to improve that is to pump it to two watts per square millimeter... It requires more complicated cold plates and complex liquid cooling, or maybe even things like immersion cooling. In space, higher watts per millimeter is very difficult, whereas on Earth, these are solved problems."

Consistent. Our satellite-gpu-capacity-scaling side page addresses thermal management as the binding constraint for concentrated satellite designs. We document that rejecting 137 kW of heat at 80C requires ~275 m^2 of radiator area, and that "thermal transport distances beyond ~10 m require mechanically pumped fluid loops with no direct flight heritage at this scale." Patel's point is that as chip power density increases (a terrestrial advantage), the thermal challenge in space becomes proportionally worse — radiative cooling scales with temperature to the fourth power but heat flux density scales linearly with power density. This reinforces our finding that thermal management is a major design driver above ~100 kW on a single satellite.

Point 16: Elon wins by 10x, not 20% — space data centers are eventually a 10x play, not this decade ↗

Patel argues: "Elon doesn't win by doing 20% gains. He never wins that way. Elon wins when he swings for the fences and does 10X gains... I think space data centers will eventually be a 10X gain as Earth's resources get more and more contentious, but that's not this decade."

Consistent. Our cost-parity-timeline reaches the same temporal conclusion: "The economic case for orbital compute depends on achieving optimistic-case assumptions and compressing the normal 7-10 year prototype-to-scale timeline for complex space systems." Even the optimistic scenario shows a persistent ~43% premium through 2035. The "eventually makes sense" framing aligns with our model, which shows the ratio declining over time as launch costs fall but never reaching parity within the analysis window (through 2040). Patel's characterization that orbital needs to be a 10x gain, not a marginal improvement, is an insightful framing our analysis does not use but that captures the economic reality.

Point 19: Big Tech capex approaching $600B/year with total supply chain at ~$1 trillion ↗

Patel states combined Big Four (Amazon, Meta, Google, Microsoft) forecasted capex this year is $600B, with the total supply chain getting to "the order of a trillion dollars." He notes Amazon can build data centers in as fast as eight months.

Consistent. These capex figures are referenced in our source summary for patel-2024-ai-bottlenecks in sources.md. The 8-month data center construction timeline is a useful benchmark against orbital deployment timelines (3-6 months for testing/integration + launch scheduling + orbital commissioning). Our analysis implicitly uses this comparison in the deployment speed discussion.

Point 21: Scale-up domain evolution — from 8 GPUs (Hopper) to 72 (Blackwell) to 144 (Rubin) ↗

Patel explains: "With Blackwell NVL72, they implemented rack-scale scale-up. That meant all seventy-two GPUs in the rack could connect to each other at terabytes a second." He also discusses Google's TPU topology (torus, thousands of chips) and notes the trend toward dragonfly topologies with hundreds or thousands of chips in the scale-up domain.

Consistent. Our inference-networking-requirements page documents this exact progression and its implications for orbital feasibility. We cite NVL72 to NVL144 expansion and note: "NVIDIA's roadmap (NVL72 -> NVL144) explicitly grows domain size." The trend toward larger scale-up domains makes the per-satellite GPU count constraint more binding: a single satellite must host an increasingly large tightly-coupled domain to compete with terrestrial inference economics.

Point 27: Compute is most valuable immediately — speed of deployment matters ↗

Patel emphasizes: "Because we're so capacity-constrained, that compute is theoretically most valuable in the first six months you have it. We're more constrained now than we will be in the future."

Consistent. Our cost-parity-timeline discusses deployment speed as a potential business case for orbital compute (faster than waiting for grid interconnection). However, Patel's point cuts both ways: if compute is most valuable immediately, the 3-6 month orbital deployment overhead (testing, deconstruction, launch, commissioning) destroys value. Our analysis captures this through the effective lifetime penalty. Combined with Point 7 (chips are the binding constraint, not facilities), Patel's argument strongly favors deploying GPUs in the fastest available terrestrial facility rather than waiting for orbital infrastructure.

Not relevant

Point 20: Anthropic at ~5 GW and OpenAI similar by end of year — inference demand is enormous

Patel estimates Anthropic needs to reach 5+ GW by end of year to serve revenue growth, with OpenAI at a similar or slightly higher level. He calculates this from Anthropic adding $4-6B/month in revenue, requiring $40B in compute spend at current margins, translating to ~4 GW of inference capacity growth.

Not relevant. This characterizes the scale and pace of AI inference demand growth but does not directly bear on the orbital vs. terrestrial competitiveness comparison. Both orbital and terrestrial operators face the same demand signal. The implication that demand is enormous strengthens the general argument for more compute infrastructure, but the question of where to deploy it is determined by cost and feasibility, which is our analysis's focus.

Point 24: Memory is 30% of Big Tech capex and supply is acutely constrained

Patel states: "Thirty percent of Big Tech's CapEx in 2026 is going towards memory." He describes HBM as having 3-4x fewer bits per wafer area than commodity DRAM, and explains that HBM4 provides 2.5 TB/s per stack vs ~128 GB/s for DDR in the same shoreline area — an order of magnitude bandwidth difference. Memory prices have tripled, smartphone volumes are projected to fall from 1.1B to 500-600M units.

Not relevant. Memory cost and supply dynamics affect GPU/accelerator pricing equally for orbital and terrestrial deployments. Both paths use the same chips with the same memory. The memory crunch does not differentially advantage or disadvantage orbital deployment. It is captured implicitly in our GPU cost per kW_IT estimates.

Point 26: International diversification of data center locations ↗

Patel notes: "Australia, Malaysia, Indonesia, and India are all places where data centers are going up at a much faster pace. But currently, over 70% of AI data centers are still in America."

Not relevant. Geographic diversification of terrestrial data centers does not directly affect the orbital vs. terrestrial cost comparison. However, it tangentially reinforces that many countries have feasible terrestrial options, weakening the "energy sovereignty" argument for orbital compute cited in our cost-parity-timeline business case discussion.

Point 28: Robots and local compute — inference demand extends beyond data centers

Patel discusses how humanoid robots would need both local compute (for real-time actions) and cloud compute (for planning and higher-level intelligence). He argues most intelligence should remain in the cloud for efficiency (batching) and capability (larger models).

Not relevant. Robotics compute demand is outside the scope of our analysis. The implication that more inference workloads will be served from centralized data centers rather than edge devices is tangentially supportive of continued data center demand growth, but does not affect the orbital-terrestrial comparison.

Point 29: Taiwan semiconductor concentration risk ↗

Patel discusses the risk of Taiwan disruption, noting it would reduce incremental compute capacity from "hundreds of gigawatts a year" to "maybe 10 gigawatts across Intel and Samsung, or 20 gigawatts." He argues this would "drastically slow US and global GDP."

Not relevant. Taiwan risk affects global semiconductor supply equally for orbital and terrestrial deployment. While a disruption could theoretically increase the value of already-deployed orbital compute, this is a geopolitical risk scenario outside our economic analysis scope.

Point 30: The Alchian-Allen effect on AI compute — higher GPU costs push toward best models ↗

Dwarkesh raises the economic concept that a fixed cost increase (more expensive GPUs) makes the ratio between high-quality and lower-quality goods more favorable for the high-quality option. Patel agrees: "We just see all of the volumes are on the best models today, all the revenue is on the best models today."

Not relevant. This economic effect operates identically in orbital and terrestrial deployments. If anything, it favors terrestrial, where the best (newest, most capable) models can run on the latest hardware with full NVLink scale-up domains, while orbital hardware is always 1-2 generations behind.

Point 31: China's semiconductor scaling and potential to outpace the West on long timelines ↗

Patel argues that on "fast timelines, the US wins; long timelines, China wins" — meaning if AI capabilities mature slowly enough, China could build an indigenized semiconductor supply chain and outscale the West. He expects China to have fully indigenized DUV by 2030 and working but not mass-produced EUV.

Not relevant. China's semiconductor trajectory does not directly affect the orbital vs. terrestrial cost comparison for Western operators. It is a geopolitical consideration outside our analysis scope.

Point 32: Anthropic and OpenAI revenue scaling implies massive return on data center investment

Patel notes that Anthropic added $4B revenue in January and $6B in February, implying the return on invested data center capital is very high. He estimates $50B of capex was deployed to generate Anthropic's $20B ARR.

Not relevant. Revenue dynamics for AI labs do not differentially affect the orbital vs. terrestrial cost comparison. Both deployment paths would serve the same revenue-generating workloads.