Hook
Two gigawatts is not a capacity figure. It is a claim on a sovereign electron budget, and the crypto market has not priced it.
Somewhere in the last quarter, Nvidia and eight Australian companies agreed to develop roughly 2GW of AI infrastructure. No GPU count. No capital figure. No named site. No disclosed power source. No published offtake. A number, a flag, and a vendor.
I have spent nine years reading infrastructure announcements the way an auditor reads footnotes, and I want to state plainly what I think before the rest of this piece takes it apart: a 2GW announcement with no megawatt-hour contract attached is the AI era's equivalent of a 2017 token sale that listed Microsoft as a strategic partner. It is a social proof artefact wearing a physics costume. The costume is convincing because the physics is real — two gigawatts really is an enormous amount of electricity — and that is exactly what makes it dangerous to reason with.
Here is what the number means if you take it literally. Two gigawatts of continuous load, run for 8,760 hours, is 17.5 terawatt-hours a year. Australia's National Electricity Market clears roughly 190 TWh annually. Total continental generation sits between 260 and 280 TWh. A single project — if built, if energised, if sustained — is a nine per cent load increase on the largest electricity market in the southern hemisphere. That is not a data centre. That is a new class of customer, and new classes of customer do not arrive without repricing everything around them.
Three on-chain markets are already pricing the consequences. All three are pricing them wrong, and they are pricing them wrong in opposite directions. Getting that right requires doing the arithmetic in public, which is what the rest of this piece does.
Context: what the announcement is, and what it conspicuously is not
Let me separate the verified from the inferred, because the source material here is thin in a very specific and revealing way.
On the record: Nvidia, eight Australian counterparties, approximately 2GW of AI infrastructure, described as a collaboration. Missing from the record: the names of the eight, the capital structure, site selection, position in the interconnection queue, power purchase agreements, GPU SKUs, network fabric, cooling architecture, the training-versus-inference split, and — the detail that determines whether any of this is investable — whether it is a binding purchase order or a memorandum of understanding.
If you have watched this asset class for more than one cycle, you know that distinction is the entire trade. An MOU has an expected value near zero and an announcement value of roughly one green candle. A binding engineering, procurement and construction contract with a capacity reservation and a parent guarantee has an expected value you can discount, hedge and finance. The gap between those two documents is where retail capital gets harvested, every cycle, without exception, and the harvest is always dressed up as a first-mover advantage.
Now the part I can reconstruct with high confidence, because I have watched four versions of this pattern.
Nvidia does not build or operate most of these facilities. It sells the rack, the fabric, the reference architecture and the software stack, and increasingly it takes an equity or revenue-share position to lubricate the transaction. The local partners supply land, capital, interconnection rights, construction and the customer relationship. The vendor's actual product is not a GPU. It is a system-level standard that makes any alternative accelerator a rewrite rather than a swap. That is a moat built out of switching costs, and it is the most successful such moat in the history of computing.
That matters for how you model the deal. When the vendor's leverage is architectural rather than contractual, the vendor does not need binding volume commitments to win. It needs the ecosystem to assume the commitments exist. That is a subtler and more durable form of lock-in than a purchase order, and it is why so many of these announcements are never converted into disclosed backlog. The backlog is the installed base, and the installed base is assumed.
In Australia, the plausible composition of an eight-company consortium writes itself. Two listed data centre operators. One or two energy majors or gentailers. A transmission or pipeline infrastructure owner. A telco for fibre backhaul. An engineering and construction firm. An asset manager to arrange the debt stack. And — the participant crypto readers should be watching — at least one bitcoin miner that has spent three years retrofitting substations, buying transformers with multi-year lead times, and learning to speak the network operator's dialect.
I will name the elephant without pretending I have inside information. The largest listed operating entity on Australian soil that has already converted mined Bitcoin revenue into contracted AI cloud revenue is the same company the market valued, as recently as 2024, as a leveraged proxy for BTC. Sydney-domiciled, United States-listed, sitting on interconnection capacity that took years and eight figures of engineering studies to secure. If this 2GW consortium has a spine, that company is a member of it, a direct competitor to it, or an acquisition target for it. Those are three different trades and only one of them is obviously good.
Why does a cross-border payments researcher care about an Australian electrical interconnection queue?
Because in every previous cycle, the marginal buyer of crypto assets was a speculator with a leveraged account. In this cycle, the marginal buyer of the physical substrate has a power bill, a credit rating and a twenty-year planning horizon. Once an asset class acquires a cost of production denominated in megawatt-hours, its price has to start behaving like a commodity and stop behaving like a lottery ticket. That transition is neither smooth nor kind to anyone still holding the old valuation framework.
The energy ledger: 17.5 terawatt-hours and the price of firmness
Let me do the arithmetic almost nobody does in public, because institutional framing is easier than institutional math.
Two gigawatts continuous is 17.5 TWh a year. That is generation-scale demand, not load-scale demand — it exceeds the annual residential consumption of a mid-sized Australian state. It has to be served by something, and the something determines everything downstream.
Australia's coal fleet is retiring on a published schedule through 2035, with the largest remaining units exiting between 2028 and 2031. The replacement volume is renewables plus storage plus firming. Renewables are cheap energy but they are not firm capacity. A 2GW hyperscale AI campus needs somewhere between 95% and 99.9% availability depending on workload mix. Training runs are partially interruptible if checkpointing is designed well; inference serving is not. So the project needs one of three power structures, and each carries a different on-chain footprint.
The first is a long-dated corporate PPA with an existing or new-build renewable portfolio, backed by four-hour batteries. In the Australian market that means Large-scale Generation Certificates attached to the offtake. LGCs are a real tradeable commodity with a forward curve, a compliance obligation and an annual true-up. They are also deeply boring, and boring commodities are exactly what tokenisers eventually get around to. The interesting question is not whether some team will put an LGC tranche on-chain. It is whether the token legally is the certificate or merely a pointer at a custodian holding the certificate. In every real-world-asset structure I have reviewed since 2021, that distinction is where the yield goes to die.
The second is self-generation. A consortium at this scale can simply build. Two gigawatts of firm capacity is roughly four to six gigawatts of nameplate solar plus a large storage fleet, or a gas portfolio with carbon accounting that will be litigated for a decade. Self-generation converts the project from an energy customer into an energy principal, which changes the financing math completely — and it is the version where tokenised project finance finally has a coherent story, because the cash flows are contracted, the assets are physical and the counterparties are named.
The third is the peaker hedge. Contract most of the energy, top up with gas turbines during Dunkelflaute events. Cheapest on paper, most exposed to gas volatility, and it produces the most interesting derivative surface because the facility's operating margin becomes a spread between GPU rental revenue and spark spreads. A GPU-hour contract, structurally, is a spark spread with extra steps: you are selling a time-sliced, location-specific, capital-intensive conversion of electricity into a service, and your margin is the difference between what the service clears at and what the electrons cost. Once you see it that way, you cannot unsee it, and the entire "AI infrastructure" category becomes legible as an energy trading business with a compute front end.
There is a fourth constraint that gets almost no airtime and it is currently the binding one: electrical equipment. Large power transformers, high-voltage switchgear, gas turbines and liquid-cooling manifolds are all on multi-year lead times, and the order books are full through the end of the decade. Two gigawatts of new load requires roughly six to ten new grid-scale transformers depending on configuration, plus substation works, plus transmission upgrades that require their own approvals and their own queues. You can order GPUs and receive them in a quarter. You cannot conjure a 500kV transformer that has not been built yet. Announcements are cheap; transformers are not.
Now connect it to crypto, because this is where the market's model breaks.
Hashprice — daily revenue per unit of hashrate — is a good leading indicator of miner health precisely because bitcoin miners are the lowest-firmness, highest-optionality buyer of power on the market. They are the grid's demand response mechanism. They buy when power is cheap and curtail when it is expensive. AI data centres do the opposite: they buy firmness, accept curtailment penalties and pay a premium for the privilege.
So when 2GW of demand bids for firmness in a 190 TWh market, the clearing price of firmness rises. What does that do to the marginal miner? It does not kill them. It pushes them further down the merit order, into the two-hour windows where nothing else wants the electron. It compresses their margin — and it also hands them a hedge they never had, because their flexibility finally has a price.
I have watched this movie in a different currency. In 2017 I was a junior quant in Istanbul modelling initial coin offering treasury velocity, four months of on-chain forensics across five hundred token sales, and the finding that stayed with me was not that sixty per cent of primary liquidity recycled within four hours. It was that apparent demand and structural demand were almost entirely uncorrelated. Announcements moved price; cash flows moved nothing. Australia's 2GW figure is being read by the market as structural. On the current evidence it is announcement. Tracing the liquidity ghosts through the ICO fog is the only skill that has ever paid me reliably, and the fog here is measured in gigawatts.
The miner multiple arbitrage, and why the beta broke
Here is the number that actually matters, in one sentence: a megawatt of AI colocation generates roughly two to three times the annual revenue of a megawatt of bitcoin mining, and sells for five to eight times the multiple.
Walk through it. A megawatt of current-generation ASICs running near 17 joules per terahash produces roughly 58 petahashes per second. At a bull-market hashprice in the mid-forties per petahash per day, that is about $2,600 a day, or roughly $950,000 a year before power. Power at a competitive industrial rate of $40 to $60 per megawatt-hour consumes $400,000 to $500,000 of that. Gross margin per megawatt per year lands around half a million dollars — and it is violently sensitive to hashprice, which is the whole problem.
Now the same megawatt as AI colocation. Contracted at $180 to $280 per kilowatt-month — and in the tightest markets we have seen prints well above that — revenue lands between $2.2M and $3.4M per megawatt per year, with three-to-ten-year take-or-pay terms, escalators, and creditworthy counterparties. Margin is higher, volatility is a fraction, and the contract duration is what makes the debt financeable at investment grade.
The multiple follows mechanically. Mining cash flows trade at three to five times EBITDA because nobody will underwrite hashprice beyond eighteen months. Contracted colocation cash flows trade at fifteen to twenty-five times because the market can underwrite a lease. Same electrons. Same substation. Same transformer that took thirty months to arrive. A completely different story.
That is why the listed miner-to-AI conversion trade has been one of the cleanest risk-adjusted expressions of this cycle, and why it is now largely crowded. The second-order effects are where crypto-native readers should be looking, and there are three that matter.
The beta broke. For a decade, listed miners were the highest-beta expression of BTC in public equity markets. As they convert to contracted AI revenue, correlation to BTC compresses and correlation to the AI capex cycle rises. If you are still using them as a leveraged BTC proxy, you are running a position you no longer understand. I have watched this exact error before: 2020, a cohort of DeFi tokens whose economic value had quietly migrated into something else while everyone kept modelling them as fee-share instruments. The model did not become wrong. The asset did.
The hashprice floor moved. When a marginal miner holds an option to convert a megawatt to colocation, the reservation price of hashrate changes. A miner will not run ASICs at negative contribution margin when the alternative is a twenty-year contracted lease with a creditworthy tenant. That should structurally raise the hashprice level at which hashrate capitulates. Anyone still running a 2019-vintage hashprice model is outputting noise with a confidence interval attached. And because hashrate is a global market, the effect propagates: capacity that leaves the bitcoin network for AI in Australia does not reappear in Texas or Paraguay at the same difficulty assumption. Network difficulty is a lagging function of a decision-making process that now includes a hyperscale lease as an alternative use of capital.
And the third one, the one I keep circling: an AI colocation contract is a tokenisable cash flow that almost nobody is tokenising, because the counterparties are boring. A take-or-pay lease with a hyperscaler, an investment-grade tenant and a fixed escalator is the platonic ideal of a real-world asset. Defined cash flows, defined termination, defined credit risk, no narrative premium. Which is precisely why it will be tokenised last and priced correctly first. The market does not tokenise what is safe. It tokenises what is exciting, and then discovers the difference during a drawdown.
DePIN compute gets repriced down at the top and up at the edge
Now the part of the market that is going to have a bad time, and I want to be precise about which part.
The decentralised compute thesis rests on three claims: idle GPU supply is abundant, aggregation is the hard part, and price competition will let decentralised networks undercut hyperscalers. The first claim is true. The third is about to become false at the top of the market and remain true at the edge.
The mechanism is simple and brutal. Two gigawatts of Nvidia-backed, debt-financed, sovereign-adjacent, politically subsidised compute does not behave like a competitive supplier. It behaves like a price setter that can run at a loss for five years because the return is strategic rather than financial. Australia is not building this because the internal rate of return on GPU rental is compelling. Australia is building this because compute is becoming a matter of state capacity, and because the alternative is dependence on three American clouds and one Taiwanese fab.
When the marginal supplier's objective function is national capability rather than return on capital, price competition stops working the way the textbook says. Decentralised networks compete on price against suppliers who do not care about price. That is a structurally losing position in the training market, and it will remain a structurally losing position for four to six years.
Where decentralised compute wins is the other end: inference at the edge, sub-100ms latency requirements, data that legally cannot leave a jurisdiction, workloads where the buyer wants verifiable execution rather than the cheapest floating-point operations. Those niches are real and growing. They are also niches, and the tokens in that sector have been priced as though they were going to take the training market.
The bear case the sector's own marketing hides is this: the product decentralised compute actually sells is not compute. It is verifiability. And verification has a cost, a latency, and a trust assumption of its own. Every time I have stress-tested a decentralised compute design, the binding constraint was never GPU availability. It was the cost of proving the GPU did what the payer believes it did. Optimistic verification with fraud proofs works for a settlement layer clearing every twelve seconds. It does not work for a thirty-hour training run whose results you need this afternoon.
There is an uncomfortably direct analogy to the oracle problem, so I will make it. A decentralised network routing compute through a small set of high-performance attestation nodes is not decentralised compute. It is centralised compute with a token attached — a consortium chain wearing a consensus mechanism as a costume. The reason the costume persists is that the alternative, genuinely verifiable general-purpose compute, is somewhere between very hard and currently impossible at competitive cost. That is not a criticism of the teams building it. It is a criticism of the valuations, which price the costume as though it were the body.
The payment rail: machine-speed settlement and the L2 cost model that cannot survive it
Here is where my day job shows up.
I have spent the last eighteen months building and stress-testing models for autonomous agent payments: wallets controlled by software rather than people, sub-cent transactions settled atomically between counterparties that have never met and never will. The consensus forecast in that space lands somewhere in the tens of billions of dollars annually by the end of the decade. I modelled $50 billion and I think that is the right order of magnitude, possibly generous on timing.
The piece nobody modelling that market has internalised is this: the transaction profile of the machine economy is the exact opposite of the transaction profile Layer 2 rollups were designed to serve.
A human user does five transactions a day at an average value of fifty dollars. An autonomous agent settling GPU time does five thousand transactions an hour at an average value of three-tenths of a cent. The economics of the second workload depend entirely on marginal cost per transaction, and the marginal cost of a rollup transaction depends almost entirely on data availability cost.
Which brings me back to a claim I have been making since the Dencun upgrade shipped and everyone celebrated: blobspace will saturate within two years, and when it does, rollup fees roughly double, and then double again. The current fee level is not a technological achievement. It is an artefact of temporary oversupply. Dencun introduced a separate data availability fee market with a target and a maximum; when aggregate rollup demand exceeded the target, the market did what separate fee markets do, and the target was later raised as a reprieve. Reprieves are not architectures. The moment demand catches the raised target — and agent payments are a demand shock arriving on roughly that schedule — the entire machine-economy cost model, built on fractions of a cent, stops working.
Now put the 2GW buildout next to that. Sovereign AI infrastructure feeds more machine transactions into more chains, not fewer. It is a demand shock for high-frequency settlement capacity arriving on a supply-constrained timetable. If you are modelling agent payments on current L2 fees, you are modelling on the cheapest quarter in the history of the technology, and you are about to be wrong in a way that is invisible until it is sudden.
There is a second, quieter mechanism worth naming. Machine payments are not just frequent; they are float-negative for the intermediary. When settlement is instant and value is atomically transferred, nobody holds the working capital. That destroys a revenue line the traditional payment industry has lived on for fifty years — the float on the two-day settlement window. Every stablecoin rail that replaces a correspondent banking leg removes a balance sheet, and balance sheets are where the incumbents actually make their money. The incumbents know this. It is why the loudest voices in tokenised deposits are the ones with the most to lose.
Where does machine-to-machine value actually settle? My honest answer is that most of it will not settle on a general-purpose chain at all. It will clear on a small number of high-throughput, near-centralised systems run by the entities that already hold the compute and the customers, with a public chain used for periodic netting and dispute resolution — the same way the current financial system uses a real-time gross settlement rail for finality and something cheaper for the ninety-nine intermediate steps. That is not a defeat for crypto. It is the same division of labour that every settlement system in history has converged on, and it is worth being early to rather than sentimental about.
The kilowatt-hour oracle problem
I want to spend a section on the piece of this that crypto keeps getting wrong, because the 2GW story makes it unavoidable.
Every few months someone announces tokenised energy, or a virtual power plant token, or a prosumer marketplace where rooftop solar and home batteries trade kilowatt-hours with the neighbours on-chain. I have read the whitepapers. I have reviewed two such designs in detail.
They all die in the same place, and it is not tokenomics. It is not liquidity. It is not regulation. It is the latency and trust profile of metering data.
Consider what an energy market actually needs. At the transmission level, frequency is managed on a sub-second basis; ancillary services markets in Australia clear in four-second and five-minute windows. At the distribution level, a virtual power plant needs interval data with settlement-grade accuracy. A residential smart meter reports every fifteen to thirty minutes in most jurisdictions, and it is calibrated by the network operator, not by the household.
So the on-chain representation of a kilowatt-hour is a claim about a physical event measured by a device you do not control, reported through an aggregator you do not audit, reconciled by a market operator you do not govern. Every hop in that chain is a place where the token can diverge from the electron, and divergence is not an edge case. It is the normal state of affairs.
Feed latency is the Achilles' heel of every attempt to put physical infrastructure on a blockchain, and energy is the worst case because the physical event and the financial claim run on different clocks. A price oracle that is one block stale costs you a liquidation. A metering oracle that is thirty minutes stale costs you a reconciliation — and reconciliation is where all the trust in the system actually lives. The industry talks about trustlessness while quietly building its most important guarantees out of monthly true-ups and dispute processes.
Which is why the only energy-adjacent assets that tokenise cleanly are the ones that already settle slowly. A power purchase agreement is a monthly settlement. An LGC is an annual compliance instrument. A capacity certificate is a quarterly true-up. Boring, slow, contractually defined, and therefore oracle-friendly. The exciting version — real-time peer-to-peer energy trading between households — is oracle-hostile and will stay in the pilot phase for another decade, funded by grants and enthusiasm. Tracing the liquidity ghosts through the ICO fog was how I built a career in 2017. In 2026 the fog has moved. It now smells like off-peak electricity and a token with a lightning bolt on the logo.
Sovereign compute nationalism and the return of capital controls — in hardware
Last analytical block. Let me zoom out to where this actually matters.
The most important macro development of the past three years is that compute has been reclassified from a commercial input into a strategic asset. Export controls, entity lists, fab subsidies, sovereign AI funds, national compute strategies: the entire apparatus of industrial policy has been rebuilt around a single scarce input in roughly thirty-six months.
Australia building two gigawatts with Nvidia is not primarily a commercial decision. It is a strategic alignment decision — the same category of act as a central bank accumulating gold, or a government filling a strategic petroleum reserve. The return is denominated in optionality and security, not in basis points.
Follow that logic to its conclusion, because the crypto market has not.
If compute is a strategic asset, then permissionless compute markets are not a neutral technological development. They are a channel for the uncontrolled movement of a strategic asset. Every time in history a valuable asset has been declared strategic, the mechanism that follows is not prohibition. It is a licensing regime, a reporting regime, a whitelist, and a grey market. The grey market is not a bug in the policy. It is the policy's shadow, and it is always priced.
I watched this exact dynamic in capital flows. In 2017, when Chinese and Korean authorities shut domestic exchange access, capital did not stop moving. It moved offshore, through intermediaries, at a spread. The spread was the price of the constraint, captured by whoever held the relationships. The identical structure is forming around compute right now, and the crypto asset that benefits is not the one with the biggest GPU marketplace. It is the one that can move value across a jurisdictional boundary without asking permission, because that is what the constrained party actually needs.
There is a second, quieter macro channel that I care about professionally. Two gigawatts of AI infrastructure is an import cycle before it is anything else. The GPUs, the networking, the liquid cooling manifolds, the switchgear — all of it is invoiced in dollars, and Australia is a mid-sized open economy funding a capital account expansion in a foreign currency. That dollar funding requirement is exactly the kind of flow that has been migrating onto stablecoin rails in the corporate-to-corporate segment for three years. Not because the treasurers are crypto enthusiasts. Because the correspondent banking chain for a mid-sized Australian importer paying a Taiwanese contract manufacturer settles in two to five days and costs real money, and a dollar stablecoin settles in seconds for a few basis points.
That is the quiet part of this story and it is the part I would actually underwrite. The loud part — tokenised GPUs, decentralised compute marketplaces, agent economies — is downstream of the boring part where a treasurer in Melbourne needs to pay a supplier in Taipei on a Friday afternoon and does not want to wait until Wednesday. Nobody writes a thread about that. It is where the volume is.
Contrarian: the decoupling thesis is running backwards, and 2GW is the new TVL
Let me now say what I actually believe, which is not what most of the market wants to hear.
The consensus narrative is that AI is eating crypto's lunch. Talent left, capital left, narrative left, and crypto is now reduced to a settlement layer for a machine economy it does not own. I think that framing is lazy, and backwards in the way that matters.
The decoupling thesis nobody is running: crypto is not being eaten by AI. Crypto is being de-financialised by AI, and that is a different and more interesting event.
For fifteen years, the primary function of a token was to be a financial claim on a narrative. Supply schedules, emissions, unlock cliffs, governance rights, fee switches — the entire apparatus exists to convert belief into a tradeable instrument. That apparatus is now being pointed at AI, and it is being pointed at assets with no narrative optionality at all. A contracted data centre lease has one cash flow and one termination clause. You cannot govern it. You cannot fork it. You cannot airdrop it, and nobody wants a governance token in a transformer yard.
When you tokenise a boring cash flow, you discover the token has no purpose beyond the cash flow. The speculation premium evaporates. A market where speculation premiums evaporate looks like the bond market: useful, enormous, and completely uninterested in your Discord server.
The 2GW number is the test case. If it becomes two gigawatts of tokenised contracted capacity with real offtake, it will be the moment the real-world asset sector stopped being a marketing category and became a market. If it becomes two gigawatts of announcements with a token attached, it will be the moment the market learned that gigawatts are a marketing unit the way total value locked was a marketing unit — and that the two have the same relationship to reality, which is a complicated one that improves dramatically when you define your terms.
Bear case, stated cleanly, three failure modes.
Failure mode one: it never gets built. An MOU between a vendor and eight counterparties with no capital committed, no interconnection agreement, no environmental approval, no offtake and no named site has an execution probability I would put below one in five over a five-year horizon. Transmission connection for loads of that scale is measured in years in every developed market. Environmental approvals for the water and land footprint are contested before they are filed. The grid does not currently have two gigawatts of firm spare capacity near a fibre route. If the announcement is the whole story, the tokens that front-ran it get cut in half and stay there for a long time, and the people who bought the narrative will describe it as an unforeseeable shock.
Failure mode two: it gets built, slowly, and the crypto proxies get the timing wrong anyway. Five to eight years to full energisation is my base case. Token markets discount on announcement and reprice on news flow, which means crypto proxies will be fully valued eighteen months in and then grind sideways for four years while the engineering firms work through approvals and transformer lead times. I have watched this pattern in infrastructure-linked tokens since 2018. The buildout is real. The trade is not the buildout.
Failure mode three: it gets built fast, and decentralised compute gets structurally squeezed. If Australia energises two gigawatts in three years, the price of contracted compute capacity in Asia-Pacific falls rather than rises, because supply arrives faster than inference demand monetises. In that world, the tokens with the most exposure to decentralised compute pricing get repriced downward by the very thing they were supposed to benefit from. This is the scenario almost nobody is modelling and it is the one I would assign the highest probability to.
Takeaway
Two gigawatts is a claim on 17.5 terawatt-hours, roughly nine per cent of Australia's largest electricity market, a capital stack in the tens of billions, and a buildout schedule measured in election cycles. It is also, right now, a sentence.
The cycle positioning question is not whether AI compute is real. It is real, it is enormous, and it will rewire the physical economy over the next decade. The question is which layer of the crypto stack holds a defensible claim on that value, and my answer is narrower than the market's: settlement rails that work when the constraint is legal rather than physical, tokenised cash flows that were boring before they were tokenised, and verification infrastructure that can be priced honestly rather than assumed away.
Everything else is an announcement with a lightning bolt on it.
Here is the single thing I am watching into the next quarter. When the eight names surface, how many of them have signed interconnection agreements, and how many have signed memoranda of understanding? The first number is the trade. The second number is the story. In every cycle I have traded, the money was made by the people who could tell those two apart before the price did.
So which is it — two gigawatts of capacity, or two gigawatts of vibes? Watch the interconnection queue, not the press release. The electrons will not lie. The tokens will not warn you.