This article is for information only. It is not investment advice. It is not a recommendation to buy or sell any security or commodity. Do your own work. Speak with a qualified adviser.
Goldman Sachs spent this week asking the right financial question.
How much revenue do the hyperscalers need to justify the next wave of AI spending?
The answer, in the latest work associated with strategist Ryan Hammond, is large.
The six names in the Phase 2 frame — Alphabet, Amazon, Meta, Microsoft, Oracle and SpaceX — are mapped against about $1.73 trillion of AI capex in 2026 and 2027.
To clear a 15% return on that capital, Goldman’s framework points to about $1.42 trillion of cumulative revenue in 2028 through 2030.
That is roughly $11.6 billion of revenue per gigawatt per year.
A related Bloomberg readout of Goldman’s work put 2027 spending by the five largest U.S. hyperscalers near $1.2 trillion. That is about 50% above this year’s estimated outlay. Break-even AI revenue sits near $300 billion a year. A “good return” for the whole stack may need closer to $1 trillion a year of end-user AI spend.
Those numbers will dominate equity desks.
They should not dominate mining desks.
The missing question is simpler.
How much metal does this take?
The city sees software. The mine sees tonnes
Most global investors live in cities.
They see tokens, cloud backlogs and GPU slides.
They do not see a haul truck, a concentrator or a smelter queue.
That is the “citiot” problem, if you want the blunt word.
It is not an insult to intelligence.
It is a description of distance.
A model can print a revenue hurdle in an afternoon.
A copper mine takes a decade.
A transformer line can take two years just to order.
A gigawatt campus cannot open on a spreadsheet.
It opens when copper busbars, aluminum conductors, silver contacts and steel racks exist in the physical world.
Goldman’s own commodities team already hinted at the scale on the power side.
U.S. data-center power demand is projected to rise from about 31 gigawatts in 2025 to 41 gigawatts in 2026 and 66 gigawatts in 2027.
Installed U.S. capacity could approach 95 gigawatts by the end of 2027 if utilization sits near 70%.
Year-over-year additions are scheduled near 13.6 gigawatts in 2026 and 36.3 gigawatts in 2027.
That is not a software release.
That is a materials order.
Start with copper, because current has to move
Copper is the metal that makes the revenue possible.
Chips do not run on narratives.
They run on current.
Current runs on copper.
Intensity estimates vary. That matters. Readers should treat every tonnes-per-megawatt figure as a range, not a law.
S&P Global has put AI training data-center copper intensity near 30 to 47 tonnes per megawatt inside the facility. Schneider-linked work has used about 39 tonnes per megawatt for AI hyperscale sites. A documented Microsoft Chicago build used more than 2,177 tonnes of copper at 198 megawatts. That is about 11 tonnes per megawatt on one published facility count, or closer to 27 tonnes per megawatt on other industry retellings of the same site.
Those gaps are not a reason to ignore the story.
They are a reason to stay honest.
The conservative reading is this.
A modern AI hall uses more copper per megawatt than a legacy cloud hall.
Power density is higher.
Busbars are thicker.
Liquid cooling adds heat-exchange metal.
Redundant feeds add more cable.
A one-gigawatt campus is often described as locking 20,000 to 50,000 tonnes of copper, depending on whether the count stops at the building or follows the wire back to the grid.
Wood Mackenzie has said the supporting power system can multiply facility metal demand by three or four times.
That is the part city models skip.
They count the rack.
They forget the substation.
S&P Global’s longer copper path is useful here.
Data-center copper demand is forecast to rise from about 1.1 million tonnes in 2025 to about 2.5 million tonnes by 2040.
AI training sites could account for 58% of data-center copper use by 2030.
JPMorgan has put incremental AI data-center copper near 110,000 tonnes in 2026 versus a pre-AI baseline.
Other houses print higher incremental numbers.
The range is wide because scope is wide.
Some counts include only the hall.
Some include generation, transmission and transformers.
Investors should ask which count they are using before they treat any headline tonne figure as destiny.
The direction is still clear.
More gigawatts means more copper.
And copper mines do not appear because a hyperscaler files a 10-K.
Now do the gigawatt math Goldman already published
Take Goldman’s U.S. addition path as a thought experiment, not a precise bill of materials.
If the United States adds about 13.6 gigawatts of data-center capacity in 2026 and 36.3 gigawatts in 2027, that is 50 gigawatts of new U.S. capacity in two years.
At 25 tonnes of facility copper per megawatt, 50 gigawatts would imply 1.25 million tonnes of copper in the halls alone.
At 40 tonnes per megawatt, the same two-year U.S. wave would imply two million tonnes.
Global refined copper supply is on the order of 26 to 28 million tonnes a year.
So even the mid-range hall-only math is not a rounding error.
Add the grid and the number gets larger again.
CIBC has used a conservative 27 tonnes per megawatt for direct data-center copper and noted that AI clusters can run 40 to 60 tonnes per megawatt.
It also pointed to more than a million tonnes a year of grid copper tied to data-center load by 2030 in Wood Mackenzie work.
That is the physical constraint hiding under Goldman’s revenue hurdle.
You cannot collect $11.6 billion per gigawatt if the gigawatt never energizes.
And the gigawatt does not energize without metal.
Silver is smaller in tonnes and larger in function
Silver will not match copper on mass.
It still sits in the stack.
The U.S. Geological Survey lists silver among the key minerals in data-center server boards and circuitry. The United States imported about 64% of the silver it consumed in the latest USGS summary cited in that work.
Silver shows up in three places.
Power contacts and plated bus work, because it resists oxidation and carries current cleanly.
Connectors and high-speed links, because contact resistance matters at data-center speeds.
Packaging, solder and thermal materials around dense chips.
Published per-server estimates are messy.
Some industry notes use 20 grams per high-performance server.
Some commercial breakdowns float 180 grams for a full AI training node when power electronics and paste are included.
Recycler estimates for a single GPU card often sit in the low single-digit grams.
Treat all of those as estimates.
Manufacturers do not publish a standard silver bill of materials.
What is not in dispute is the direction.
More racks mean more contacts.
More 800-gigabit and 1.6-terabit optical modules mean more specialty loadings. One Chinese industry note has put silver use near one gram per 800-gigabit module, several times a 100-gigabit unit.
Silver is also the metal of the solar panels that some campuses use to decorate their power story.
That last line is easy to oversell.
Data-center silver is not yet the main pillar of the silver market.
Photovoltaics and electronics still dominate industrial demand.
The point is narrower.
The same investors who debate token prices rarely ask whether the contact metal is in surplus.
It is not a large market.
A few tens of millions of extra ounces matter.
Aluminum and steel do the unglamorous work
Aluminum moves heat and long-distance power.
It sits in heat sinks, cold plates, racks and the overhead lines that feed a campus.
Wood Mackenzie has said cooling can account for about 55% of in-facility aluminum use, with racking another 25%.
Transmission is where aluminum really scales.
Steel still dominates the mass of a large site.
One widely circulated materials breakdown put a one-gigawatt AI campus near 180,000 tonnes of raw materials, with steel around three-quarters of that mass and copper a distant second.
Tin matters in solder.
An AI server can use several times the tin of a conventional box.
Tantalum, palladium, gold and gallium show up in smaller amounts.
The USGS import scores on several of those inputs are high.
This is not a “copper only” story.
It is a bill of materials story.
City investors talk about models.
Mines talk about concentrates.
Those are not the same language.
What Goldman left out of the denominator
The ZeroHedge write-up of the Goldman work made a second financial point that miners should notice.
A large share of the build is not sitting cleanly on the hyperscaler balance sheet.
Leases, special-purpose vehicles and off-balance-sheet structures can hide part of the true capital base.
If the denominator is understated, the advertised return looks better than the project return.
That is a finance problem.
It is also a materials problem.
Metal still has to be bought whether the lease sits on or off the books.
The transformer still has a lead time.
The copper still has to be mined, smelted and drawn.
Morgan Stanley’s more skeptical work, cited in the same debate, has asked whether supply and rental prices can keep justifying the capex.
Token prices have already shown they can collapse.
That is the bear case for the software layer.
It is not automatically a bear case for the metal layer.
If the build slows, metal demand slows with it.
If the build continues as an arms race, metal demand does not need a pretty ROIC to stay real.
Governments and platforms can fund ugly returns for a long time when the prize is compute.
They cannot fund a mine that does not exist.
The backlog is paper. The wire is not
Goldman points to cloud backlogs as the demand offset.
AWS, Azure and Google Cloud combined backlog was cited near $1.69 trillion in the second quarter of 2026, up about 152% year over year.
On Goldman’s own math, a large slice of 2026–2027 capex could be covered by that backlog if it converts.
Backlog is a claim on future revenue.
It is not a claim on future cathode.
Conversion still needs power.
Power still needs metal.
Grid interconnection queues in parts of the United States now stretch years.
Transformer delivery times have been reported beyond two years.
That is why some campuses look for behind-the-meter generation.
Gas turbines, nuclear restarts and on-site plants are not software either.
They are more steel, more copper and more time.
Why urban investors keep missing the constraint
Software scales with copy and paste.
Mines do not.
That is the whole educational point.
A model can raise capex by $200 billion in a cell.
A copper project cannot.
Permitting, water, community consent, smelter capacity and grade decline all sit between the spreadsheet and the busbar.
Recycled copper helps.
It does not fully replace new mine supply in high-spec electrical uses, and it does not appear on the schedule a hyperscaler wants.
Aluminum can substitute in some conductors.
It cannot substitute everywhere conductivity and space are tight.
Silver can be thrifted in some pastes and contacts.
It is hard to thrift to zero when reliability is the product.
City investors are trained to underwrite code.
They are not trained to underwrite ore.
That is why the Goldman note feels complete in a Manhattan conference room and incomplete in a truck shop in the Highland Valley or the Atacama.
The Canadian layer is the physical option
Canada does not have to win the model-training race to matter here.
It has to remain a place that can deliver metal into a tight market.
That is a different contest.
Teck Resources is the large Canadian name most often cited when investors want copper torque. Second-quarter 2026 copper output was reported at 135,900 tonnes, up 25% year over year, with a sharp jump in adjusted EBITDA. That is an operating fact, not a recommendation.
First Quantum, Hudbay, Lundin Mining, Capstone and Ivanhoe sit in the same wider copper conversation, each with different jurisdictional mixes and project risk.
Silver exposure in Canada is often a by-product story.
That still matters if industrial loadings keep rising while the silver market stays small.
None of these names is “the AI trade” in the way Nvidia is the AI trade.
They are the trade underneath the AI trade.
If the buildout is real, they sell a scarce input.
If the buildout is a bubble, they still sell a metal the grid needs for EVs, transmission and ordinary industry.
That dual use is the investor’s margin of safety.
It is also why the metal question is more durable than the token question.
How to think about intensity without fooling yourself
Use three buckets.
Bucket one is the hall.
Servers, busbars, cooling plates, switchgear.
This is the 20 to 50 tonnes of copper per megawatt zone, with published site examples that sit lower.
Bucket two is the wire to the fence.
Substations, transformers, redundant feeds.
This is where intensity jumps.
Bucket three is the system.
New generation, long transmission, and the gas or nuclear plant that keeps the campus alive at 2 a.m.
Wood Mackenzie’s multiplier lives here.
Investors who only model bucket one will understate metal.
Investors who treat the most aggressive 50,000-tonnes-per-gigawatt headlines as a law will overstate it.
The adult posture is a range, updated as more as-built sites are disclosed.
Microsoft’s Chicago figure is useful because it is a real building, not a tweet.
S&P’s 30 to 47 tonnes per megawatt is useful because it is an engineered intensity, not a slogan.
Goldman’s gigawatt path is useful because it is a demand schedule, not a price target.
Put those three together and you have a framework.
You do not have a sure thing.
The opportunity is not “AI stocks.” It is scarce physical supply
The central idea is not that every copper equity must rise.
The central idea is that the AI capex cycle is a materials cycle wearing a software costume.
If hyperscalers spend $1.2 trillion to $1.7 trillion in a tight two-year window, they are bidding for the same finite streams of copper, aluminum, silver, tin and transformer steel as the grid and the auto industry.
They are price-insensitive relative to a miner.
Copper is a rounding error in a $40 billion campus.
It is the whole business of a mid-tier producer.
That asymmetry is the opportunity.
It is also the risk.
If token prices stay weak and rental rates fall, some campuses slip.
If power cannot be interconnected, some campuses slip.
If both happen, metal demand from this one end use cools.
The hedge is that copper and silver are not only AI metals.
They are electrification metals.
The AI wave is an extra call on a market that was already tight.
What a serious investor watches next
Watch announced gigawatts that actually interconnect, not announced gigawatts on a slide.
Watch transformer lead times.
Watch treatment charges and smelter availability.
Watch whether aluminum substitution shows up in busbar specs.
Watch silver industrial demand in electronics and solar, not just the data-center footnote.
Watch Canadian permitting speed, because a mine that cannot be built cannot fill a deficit.
Watch the difference between hall copper and grid copper in every new research note.
If a bank says “AI needs X tonnes” and does not define the fence line, discard the precision and keep the direction.
People also asked
How much copper does an AI data center use?
Published ranges for the facility itself often sit between about 20 and 47 tonnes per megawatt. A one-gigawatt campus is commonly described as 20,000 to 50,000 tonnes of copper once power gear is included. Grid buildout can multiply that again. Treat every exact tonne count as an estimate.
Did Goldman calculate the metal required for $1.7 trillion of AI capex?
The September 2026 Goldman work that hit the tape this week is a revenue and return framework. It asks how much money the spend must earn. Separate Goldman commodities work has mapped the power demand in gigawatts. The metal conversion is the step most equity notes still skip.
Is silver a real AI metal?
Yes, in small loadings per server and module. No, it is not the mass metal of the build. Copper and aluminum dominate tonnes. Silver matters because the market is smaller and the functions are hard to replace.
Could the buildout fail and still leave miners useful?
Yes. A slower AI cycle would cut one demand slice. It would not cancel grid spending, vehicle electrification or the long copper supply gap S&P still sketches into 2040. That is a scenario, not a promise.
The sentence Goldman did not write
Goldman can tell you how much revenue a gigawatt must earn.
It cannot tell you the gigawatt exists until the metal does.
$1.42 trillion of 2028–2030 revenue is a model output.
Tens of thousands of tonnes of copper per gigawatt are a mine output.
Those two facts do not live on the same floor of the building.
That is why so many urban investors can recite ROIC and still not know how much metal the boom consumes.
The opportunity for a resource reader is not to win a software argument.
It is to own, or at least understand, the physical layer the software argument assumes.
Money can be printed.
Tokens can be copied.
Copper cannot.
This article is for informational and educational purposes only. It is not investment, tax or legal advice. It does not recommend any security, commodity or course of action. Company names are used as industry examples. Past performance and published forecasts are not guarantees. Mining and commodity investing can result in the loss of some or all capital. Canadian Mining Report and its contributors may hold positions in securities mentioned from time to time. Verify all figures with primary filings and current market data before acting.