22. Is Taking Tokens Overseas Cyber Power Theft? Why Can't the U.S. Build Its Grid? — Ian | Power Guy

Summary

• Ian’s take on taking tokens overseas is a bucket of cold water: power shortages at data centers may account for only 30% to under 50% of the token price, or even less. He stresses that infrastructure, cooling and other costs make up the balance, while Raymond adds differences in model vendors’ quotes, labor, chips and inference costs—“exporting power can only be a gimmick.” The backdrop is OpenClaw, the “little lobster” agent that propelled Kimi, MiniMax and GLM; MiniMax gained 30%-40% after listing on the back of the trade. Conceptually, though, this is no different from exporting steel: use cheap domestic power to make something and sell it abroad.
• On fears that AI data centers will overwhelm the North American grid, Ian offers the episode’s most tradable line: “Blackouts are unlikely, but higher power prices are 100% certain.” A broad, prolonged collapse is unlikely, and Ian says outages today are caused by wildfires, blizzards and other natural events. But as long as data centers stay connected to the grid for backup power, the investment enters the rate base and is ultimately spread across every customer; higher residential power bills are “unavoidable.”
• Ian attributes the aging U.S. grid to old equipment, a conservative path from technology to deployment, and weak demand and institutions: most assets were built in the 1960s-1980s, electricity demand grew only 1%-2% a year over the past 20 years, and “without demand, there is no growth”; Raymond questions whether that figure should actually be negative, while Ian says the number may not be exact. Add the fragmented ISO/RTO structure and strong states against a weak federal government, and FERC cannot compel transmission construction under the Federal Power Act—leaving the power system divided among local fiefdoms. Sending Washington hydropower to New York would require crossing roughly 11 states, and any one of them could block the line; local generators could also influence ISO votes. “The outsider monk can’t even preach here.”
• AIDC’s power profile is a shock the grid has barely seen, while cost attribution has a blind spot around reliability: GPU load can “go from 0% to 100% in a second,” training starts at 10,000 GPUs and runs for five days before dropping to zero, yet reliability investment is handled through the ancillary-services market and “the data center pays a cent.” Bitcoin mines are different: they are highly sensitive to power prices and can shut down at any time, “a balancing tool for the grid rather than a shock.” That is why converting a mining-site power shell into an AIDC now requires a fresh grid-feasibility analysis.
• Four- to seven-year interconnection queues are being bypassed along two routes: connecting as interruptible load, so the data center gets cut first in an emergency—sometimes by several GW—and using behind-the-meter generation while seeking permitting room in selected states. xAI has put methane gas turbines on trucks in an effort to classify them as non-primary generating facilities, while Raymond cites GEV as an example of interstate regulatory arbitrage. Hyperscalers are also joining queues in 5-6 states and ultimately building in only one—“like registering at 10 hospitals and choosing one after you get called.”
• Storage is not the bottleneck holding U.S. wind and solar back; cost, policy and resource endowment matter more. The U.S. imposes a 200% tariff on Chinese solar panels, and many projects only work with the 30% ITC credit; “Trump came in and cut the ITC—the 30% share is too large for power prices alone to carry.” The U.S. and Canada also have ample natural gas and less need for new energy sources, while China has a stronger energy-independence imperative and government support for renewables. Nuclear is an option AIDC cannot wait for: Microsoft bought an existing plant for life extension, as have most companies in similar deals, while a greenfield build takes 10 years—far too slow for data centers targeting 1-2-year commissioning.
• One highly time-sensitive but uncertain signal: Ian says the Stargate Texas project had planned to expand from 1.2 GW to 2.1 GW but “can’t expand now because of funding,” and suggests Oracle’s recent stock move may also be related, while acknowledging he does not know enough about it. Raymond’s market footnote: power is a very tight balance, and whoever controls a power shell—an already energized site—gets a premium valuation. That is why Bitcoin mining sites were collectively re-rated in the last cycle.

Deep dive

1. “Power exports can only be a gimmick”: Electricity may account for 30%-50% or less of token prices

• The episode opens with the hottest narrative on the Chinese internet: OpenClaw, the “little lobster” agent, has sent token consumption soaring. Overseas users frustrated by expensive U.S. tokens have shifted to Kimi, MiniMax and GLM, while MiniMax shares jumped 30%-40% after its listing. The central question is whether low-cost domestic power can be exported in token form.
• Ian admits he is riding the trend, but reaches the opposite conclusion: data-center power shortages may account for only 30% to under 50% of the token price, or even less. Infrastructure, cooling and other costs make up the balance, and those costs are already lower in China. “Exporting power can only be a gimmick.” His analogy is steel: another power-intensive product made with cheap domestic electricity and sold overseas. Tokens simply make the dynamic more visible because they are riding the data-center wave.
• Raymond adds model vendors’ quoted prices, wages and chip-cost differences between China and overseas, particularly on the inference side. “This isn’t just about stealing electricity; it’s about the overall cost structure.” Rapid token-output growth and improving gross margins are what make the market look promising.

2. The 1960s-1980s skeleton: Why the U.S. grid is old

• Ian gives 3 reasons. First is asset age: most U.S. transmission lines and substations were built in the 1960s-1980s; wooden poles may already be 40-50 years old against a theoretical life of 60-70 years. Utilities rely on capital-maintenance teams to monitor individual assets and replace them as they expire, rather than deploying one large pool of funding to rebuild everything at once as China does. Second is a conservative path from university ideas and papers to infrastructure deployment; the U.S. may lead in research but lags China in implementation. The core reason is demand: electricity consumption grew only 1%-2% a year over the past 20 years. “Without demand, there is no growth.”
• Raymond questions that figure, arguing that after decades of deindustrialization and widespread factory closures, electricity growth “should be negative.” Maintaining 1%-2% would imply residential demand filled the gap. Ian concedes the number may not be exact but says the conclusion stands: “If the market isn’t growing, I won’t invest as a power company.” Without a business case proving demand, nobody invests.

3. Power fiefdoms: 11 states FERC cannot control

• The structural problem is clear: ISO/RTO regions run generation through wholesale-market auctions, while non-ISO regions remain vertically integrated monopolies. States are strong and the federal government is weak; FERC can issue recommendations, but “under the Federal Power Act, it cannot compel the construction of transmission lines.”
• The example carries the argument. Washington has abundant hydropower, while New York relies heavily on imported natural gas. In theory, moving power from west to east could lower New York’s electricity bills, but a line would have to cross roughly 11 states. If any one state refuses, the project cannot be built. Local generators whose interests would be hurt also hold substantial influence inside the ISO and can sway the vote. “Getting someone whose interests are hurt to vote for you is very difficult.” Ian endorses Raymond’s summary: the outsider is not merely competing locally; “the outsider monk can’t even preach here.” On power, the U.S. is almost a country of warlords.
• Two more layers of friction compound the problem. AC power cannot travel very far; beyond a certain distance, more expensive DC transmission becomes necessary. States must also split the investment, leading to endless arguments over who pays and who benefits. Elections every 4 years can reverse policy by 180 degrees, which is why headlines often report that a line has been stalled for 10 years. Innovation therefore tends to happen in relatively small, relatively independent jurisdictions such as Texas.

4. Japan is stuck on frequency; Europe is more tightly integrated than the U.S.

• Japan’s constraint is physical: Kanto runs at 50 Hz and Kansai at 60 Hz. Moving power between the regions requires DC transmission and substantial capital. The U.S. and China each use a single frequency nationwide.
• Europe’s renewable penetration has risen relatively quickly. Renewables are “both a blessing and a curse”: they are needed for decarbonization, but their intermittency makes the grid less stable. Europe therefore has to interconnect national grids so countries can support one another.
• Raymond points to the paradox: sovereign countries should be harder to connect than U.S. states. Ian’s answer is that this is “a very strange phenomenon.” Europe is even more interconnected because it depends more heavily on its neighbors. Many of Europe’s smaller countries cannot generate enough power on their own and must rely on countries with sun or wind.

5. Broad grid collapse is unlikely; higher prices are certain

• On claims that AIDC will overwhelm the North American grid, Ian first defines and then rejects the scenario. A broad, prolonged collapse is “unlikely.” Local outages—California wildfires burning transmission lines, or Texas winter storms leaving supply short and forcing rolling blackouts—happen everywhere in the world. He says outages today are caused by natural events. The core judgment: “Blackouts are unlikely, but higher power prices are 100% certain.”
• The price mechanism is straightforward. A fully off-grid, self-sufficient data center would not affect the grid, but achieving 99.99% reliability requires extensive redundant generation. The more economical option is to connect to the grid for backup, and “once you connect to the grid, you affect power prices.” Ian’s analogy: one Costco outside your home does not change traffic, but 10 Costcos require road expansion, and that cost goes into the rate base and gets allocated across every customer.
• Raymond identifies the political flashpoint: “These machines don’t eat, they don’t order delivery, and now my power bill goes up.” Ian adds that Canada is planning a data-center tax: projects above a certain MW threshold would pay 2% to upgrade the grid. It would be “three birds with one stone,” but the policy has not been implemented and the exact tax base remains unclear.

6. A power profile the grid has never seen: Cost attribution hits a blind spot, while miners provide flexibility

• Ian’s cost-attribution principle is “whoever causes it pays.” Transmission lines are relatively easy to quantify; a line built specifically for a data center should be funded by that data center. Reliability shocks are harder to measure. GPU load is not a flat line like air-conditioning demand—it can “go from 0% to 100% in a second.” No grid-connected device has historically consumed power this way, with perhaps only a small number of steel mills as a partial comparison. The equipment needed to stabilize voltage is procured through the ancillary-services market, and “the data center pays a cent.” Raymond’s verdict: “That’s not fair.”
• Raymond cites what he believes was “some kind of guidance from FERC” saying that grid-expansion costs should be borne entirely by AIDC. Ian distinguishes the reliability equipment: that investment is handled through the ancillary-services market rather than directly assigned to the data center.
• The volatility comes from training, not inference. Inference is growing rapidly but is relatively smooth; training starts at 10,000 GPUs, can run continuously for 5 days, then goes from 100% to zero. Ian admits he does not understand why everyone has to train at once instead of staggering workloads. Raymond says the reality of the past 2 years has been: “When you run, everyone runs together—brute force works miracles.” He also points to innovation in off-peak training of smaller models and intermittent training.
• That distinction changes how mining-site conversions should be assessed. Bitcoin miners are sensitive to power prices; if mining is uneconomic, they shut down and try again tomorrow. “Bitcoin miners are a balancing tool for the grid, not a shock.” AIDC training is a firm requirement: even if power prices rise, the run has to finish. That is why U.S. energy-commission guidance requires a new grid-feasibility analysis even when an existing mining-site power shell is being reused. Do not assume it is merely a change of skin.

7. How to bypass a 4-7-year queue: Interruptible load, gas turbines on trucks and 10 hospital registrations

• North America has 4-7-year interconnection queues, with Virginia hosting many data centers and facing some of the longest waits. Policy discussions from late 2025 to early 2026 point to a potentially workable route: connect as interruptible load. When power is short, “cut them first”—hospitals and schools stay on, while the data center is disconnected. Cutting several GW would immediately relieve the grid. Raymond translates the arrangement as treating a data center like “a battery that can be taken offline.” Combined with behind-the-meter generation and self-built plants, a project might connect in 3 years, while data-center developers want to be online within 1-2 years.
• Going off-grid is not a universal solution. xAI has considered methane generators, but the carbon emissions are extreme and permitting varies by state; a single strict state could make every unit unusable. Some states may offer a permitting loophole, such as mounting gas turbines on trucks and claiming they are not primary generating facilities. Raymond cites GEV as an example of interstate arbitrage: go to State A if it allows the project, skip State B if it does not. Traditional data centers are concentrated in Virginia, inside PJM; today’s hyperscale facilities are concentrated in Texas, followed by California. Ian says he has not distilled the fundamental reason for Texas, while Raymond traces it to Bitcoin miners’ migration and the conversion of mining sites.
• The more disruptive tactic is to enter interconnection queues in 5-6 states simultaneously, then choose one and withdraw from the rest—“like registering at 10 hospitals and picking one after you get called.” Ian notes that this also creates a grid problem: a substation may be built and then sit unused if the data center does not arrive, while the cost is still shared by everyone. “That isn’t fair to the local community.”

8. Storage is not the bottleneck, nuclear is too slow, and the historical burden arrives 30 years later

• Ian rejects the idea that storage is the bottleneck behind the U.S. lag in wind and solar, citing 2 main reasons. The first is cost and policy: the U.S. imposes a 200% tariff on Chinese solar panels, and many projects only work because of the 30% ITC credit—invest 20 million and receive 6 million back. “Trump came in and cut the ITC. The 30% share is too large; power prices alone can’t carry it.” The second is resource endowment: the U.S. and Canada have ample natural gas and less need for renewables to support incremental power demand. China faces a stronger energy-independence imperative, while subsidies and government direction support renewable buildout. Pairing storage with wind and solar mainly improves project returns: the U.S. and Canada have energy and capacity markets, allowing storage to buy low and sell high into power-price volatility.
• Nuclear is a long-term answer, not an AIDC answer. Most companies are buying existing nuclear plants for life extension, as Microsoft did. Ian says he has not seen a greenfield project already delivered to supply AIDC; a 10-year construction cycle is too slow for a data center but not too slow for the U.S. grid. Over the long run, nuclear would be “enormous” for energy security, generation and grid stability, provided successive administrations reach consensus and allow projects to complete roughly 10 years of permitting and construction—an extremely difficult condition. On a time-sensitive note, Ian says Stargate’s Texas project had planned to expand from 1.2 GW to 2.1 GW, but news this month says the expansion cannot proceed because of funding. He suggests Oracle’s recent stock move may also be related, while acknowledging that he does not know enough about it.
• The closing comparison is the most revealing. Canada produces natural gas but lacks pricing power; when gas prices rose in 2020, electricity prices surged as well. “A complete energy structure is equivalent to solving energy security.” Raymond’s historical view is that America’s bottom-up, state-level system may look elegant at first, but “after 30 years, it becomes a historical burden.” China is still early enough that its own burdens are not yet visible. “Our burdens may show up 30 years from now. Whether the next generation has better wisdom to deal with them will be up to them.”