What is the connection between AI chips and data center real estate? The connection is direct: every advanced AI chip needs a powered, cooled, and connected building to run in, so the supply and price of AI chips drive demand for data center space. On June 18, 2026, that link moved to the center of the market when Amazon CEO Andy Jassy signaled that Amazon Web Services may begin selling its custom Trainium AI chips to outside data centers, a business he has valued at roughly $50 billion in annual revenue. For commercial real estate investors tracking the AI infrastructure boom, this is a structural signal worth reading closely. For the broader landscape, see our guide to the best AI commercial real estate tools and strategy.
Key Takeaways
- Amazon may sell its Trainium AI chips to third party data centers, a move CEO Andy Jassy frames as a roughly $50 billion annual opportunity, up from about $20 billion in current internal chip revenue.
- Merchant AI silicon from Amazon and Google lowers the cost and widens the supply of compute, which expands the pool of operators able to build competitive AI data centers.
- For CRE investors, cheaper and more available chips support sustained demand for power rich data center space rather than signaling a near term slowdown.
- The real bottleneck is shifting from chips to power and land, with Northern Virginia vacancy below 5% and high density racks of 30 kW and above in short supply.
- Data center remains a standout asset class, with US colocation projected to grow from about $46.84 billion in 2026 to $72.37 billion by 2030.
Why AI Chips and Data Center Real Estate Are Now Inseparable
For three years the AI buildout has been bottlenecked by one company: Nvidia, whose data center division generated more than $193 billion in revenue last fiscal year and which sits on a revenue run rate above $326 billion. Amazon's move changes the supply picture. In his April 2026 shareholder letter, Andy Jassy put Amazon's annualized chip revenue at roughly $20 billion across the Trainium, Graviton, and Nitro lines, and said that if Amazon sold those chips directly to third parties the run rate could approach $50 billion. In his words, demand is strong enough that Amazon may sell racks of them to outside parties in the future.
This follows Alphabet, which in April 2026 began making its Tensor Processing Units available to external data centers through a limited customer program. Two of the three largest cloud providers are now positioning their homegrown silicon as merchant products. Trainium2 already offers roughly 30 percent better price performance than comparable GPU alternatives and has largely sold out. Trainium3, which began shipping in early 2026, adds a further 30 to 40 percent improvement and is nearly fully subscribed, with companies such as Uber moving workloads onto it. Amazon said in April that the Trainium line has generated more than $225 billion in revenue commitments. When the cost per unit of AI compute falls, the economics of building and leasing data center space improve for a wider set of operators. That is a real estate story as much as a semiconductor story.
From Chip Scarcity to Power Scarcity: The Real CRE Bottleneck
The most important takeaway for CRE investors is where the constraint moves next. If merchant silicon eases the chip shortage, the binding limit on AI capacity becomes power, land, and cooling, all of which are physical real estate problems. Northern Virginia remains the world's highest density colocation market, with vacancy below 5 percent, and high density racks of 30 kW and above are in short supply across Northern Virginia, Silicon Valley, and Chicago. Liquid cooling, including direct to chip and immersion systems, is accelerating because legacy buildings cannot dissipate the heat that modern accelerators produce. Power shortages are already pushing development toward secondary markets where substation capacity and land are available.
That is why grid access has become a defining variable for site selection and underwriting, a theme we cover in detail in our analysis of federal data center grid rules and the POWER Up Act. More available chips do not relax the power constraint; if anything, broader access to affordable silicon intensifies competition for the scarce inputs that data center landlords control. Office construction spending, by contrast, fell roughly 9 percent year over year to about $46 billion in March 2026, even as data center construction surged past it, a clear signal of where capital is flowing.
What Merchant Silicon Means for Data Center CRE Investors
- Demand durability: Cheaper compute historically expands usage rather than shrinking it. Lower chip costs tend to pull more AI workloads into production, which sustains absorption of powered shell and turnkey data center space.
- A wider tenant base: If AWS and Google sell chips externally, neoclouds, enterprises, and sovereign operators can build credible AI capacity without waiting in Nvidia's allocation queue. A broader tenant pool improves leasing depth and tenant credit diversification for landlords.
- Pricing power on power: When silicon is no longer the scarce input, the operators that control interconnection, substation capacity, and water for cooling capture the margin. Net operating income, defined as gross revenue minus operating expenses, increasingly tracks access to power rather than access to chips.
- Cap rate context: Stabilized hyperscale data center assets have traded at compressed cap rates relative to traditional commercial property because of long lease terms and investment grade tenants. Sustained chip supply supports the income assumptions underpinning those valuations.
- Concentration risk: A single hyperscaler selling chips, leasing space, and operating clouds raises counterparty concentration questions that careful investors should underwrite, not ignore.
For deeper demand context, our breakdown of Oracle's $638 billion AI backlog shows how committed cloud revenue translates into multi year leasing pipelines for data center landlords. CRE investors looking for hands on AI implementation support can reach out to Avi Hacker, J.D. at The AI Consulting Network to translate these macro signals into asset level underwriting.
The Bubble Question: Does Cheaper Compute Mean Oversupply?
The natural worry is that abundant chips plus aggressive construction equals oversupply. That risk is real and asset specific. The market is bifurcating between modern, power dense, liquid cooled facilities and older builds that cannot serve GPU intensive workloads. We work through this supply and demand balance in our piece on AI data center oversupply, and the conclusion holds here: the risk is not uniform across the asset class. Newer Class A data centers in power constrained markets remain undersupplied, while speculative product in weak power markets carries genuine lease up risk. Merchant silicon does not erase that distinction; it sharpens it by making the quality and location of the physical building the deciding factor.
The macro backdrop still favors the sector. The market for AI in real estate is projected to reach $1.3 trillion by 2030 at a 33.9 percent compound annual growth rate, and US colocation alone is forecast to grow from roughly $46.84 billion in 2026 to $72.37 billion by 2030, led by Equinix and Digital Realty. According to CBRE research, data center demand continues to outpace deliverable supply in the top US markets, with Digital Realty reporting its Americas pipeline at 79 percent leased. If you are ready to transform your acquisition process with AI, The AI Consulting Network specializes in exactly this kind of infrastructure aware underwriting. The practical playbook is to underwrite power rather than square footage, favor markets with a path to incremental megawatts, stress test tenant concentration, and treat cooling capability as a capital expenditure that protects future rent.
Frequently Asked Questions
Q: Will Amazon actually sell its AI chips to other data centers?
A: As of June 18, 2026, it remains an early stage signal, not a finalized program. Andy Jassy has said demand is strong enough that selling racks of Trainium chips to third parties is quite possible, valuing the opportunity near $50 billion. Google already runs a limited external program for its TPUs, signaling the direction of travel for merchant AI silicon.
Q: Why would cheaper AI chips be good for data center real estate?
A: Lower compute costs typically expand AI usage rather than reduce it, which sustains demand for powered, cooled, and connected space. Cheaper chips also let more operators build competitive capacity, widening the tenant base while power and land remain firmly in real estate's control.
Q: What is the biggest risk this creates for CRE investors?
A: Oversupply in the wrong locations. Abundant chips plus aggressive construction can leave speculative facilities in power weak markets struggling to lease, while modern, high density assets in power rich markets stay tight. Location, power access, and cooling capability separate winners from losers.
Q: How should data center underwriting change because of merchant silicon?
A: Shift the underwriting focus from chip availability to power, land, cooling, and tenant concentration. Confirm contracted megawatts and substation timelines, value liquid cooling capability as obsolescence protection, and diversify away from single hyperscaler dependence where possible.