The most misunderstood line item in the AI buildout is depreciation. Every hyperscaler is reporting compressed margins as tens of billions of dollars of AI capex hit the income statement through D&A while revenue is still ramping.
Amazon’s own accounting illustrates it: roughly $220B of AWS-attributable 2027 capex, 70% in short-life assets (AI servers on ~5-year lives, general servers on 6), produces about $30.6B of annual depreciation that lands in full from day one, while incremental revenue takes about three years to reach 100% utilization. Year one looks breakeven-to-negative even when the underlying investment is excellent.
SpaceXAI — the entity formed by the February 2026 SpaceX–xAI merger, public on Nasdaq as SPCX since its record $85B IPO on June 12 — is running a different version of this model, and the difference is the whole story.
The SpaceX earth based data centers projections I have made are being debated by the Limiting Factor and Jo Bkahdi.
I think a $52B/GW for new 2027 leases is defensible as a base only for capacity signed while supply stays tight. Also, $52B/GW is not even a full bull case where new Rubin chips can rent for more.
Anthropic Fable is using bear $38B (rates mean-revert toward Anthropic levels as hyperscaler supply lands), base $46B (modest haircut on $52B for volume + the per-GW plateau effect), bull $55B (Google rate holds, scarcity persists).
It only gives another 2 Gigawatts of extra power and chips (beyond 2 GW of what is already available in power and soon chips at Colossus 1 and Colossus 2) as the 2027 exit. The share price impacts can happen 4-6 months in advance when the deals are announced and projected.
Epoch.ai is indicating 440,000 chips are available and installed at Colossus 2 as of July 1, 2026. Only 223,000 of those chips are rented to Google (110k) , Anthropic (95K) and Reflection AI (18k). Current 946 MW supports the existing 440k chips. Scaling to ~2 GW IT [already almost done power wise] (with additional cooling and on-site power infrastructure already advancing) implies room for roughly another 400k–500k+ B300-class chips (depending on exact rack density, networking overhead, and PUE improvements. There will be 200k-250k chips ready to rent or announce before the end of this year. There is also a good chance of standing up the full 4 gigawatts earlier in 2027 by expanding at current locations or at new buildings.
Wholesale leasing collapses the revenue lag
AWS sells cloud services at retail through a multi-year enterprise adoption curve. SpaceXAI signs wholesale GPU leases that pay full freight almost immediately.
Three deals already disclosed:
• Anthropic: $1.25B/month (~$15B/year) for the full capacity of Colossus 1 — roughly
230,000 GPUs (H100/H200 plus early Blackwell) and 300+ MW — running through
May 2029, with 90-day mutual termination rights.
• Google: $920M/month (~$11.0B/year) for ~110,000 newer-generation GPUs, ramping
at reduced fees June–September 2026 and hitting the full rate October 1, 2026,
through June 2029. Google can terminate if committed GPUs aren’t delivered by
September 30, 2026; after December 31, 2026, either party can exit on 90 days’ notice. Google has described the deal as bridge capacity for Gemini Enterprise demand.
• Reflection AI: $150M/month (~$1.8B/year) for GB300 capacity — the third disclosed
deal, and the clearest evidence that demand extends past the two anchor
hyperscalers.
That is $27.8B/year of contracted run-rate — against an estimated ~$15B of annual
depreciation on the assets in service. The buildings are built, the turbines are bought, the chips are racked. The major capex behind these contracts is already spent. Under straight-line accounting, the leases flip the same facilities that produced 2025’s $6.4B operating loss (on $3.2B of revenue) into GAAP-operating-positive territory — on signed paper alone, before any new deal.
The pricing curve: chip generation sets the rate
The three contracts also give us something rare: an observable price curve. Anthropic’s Hopper-heavy Colossus 1 implies roughly $38–44B per GW per year. Google’s newer-mix block implies roughly $50–55B per GW per year. One generation of chip improvement is worth about 30% more per GPU per month. New 2027 capacity will be B300 (Blackwell Ultra) and early Rubin — which supports a blended new-lease assumption near $46–52B/GW/year if supply stays tight. The honest caveat per-GW rates may plateau even as chips improve, because newer chips draw proportionally more power. What improves for the tenant is tokens per dollar, which supports the same rate per GW rather than an ever-rising one.
The power ramp is the gating physical constraint
The capacity story (per industry trackers) roughly 1.75 GW on-site today, with 6–7 mobile turbines (~220 MW) added monthly through a locked supply chain — a 49.9%-owned Solaris JV, an APR Energy acquisition, 60–65% of Solar Turbines’ Titan production, and twelve 380 MW Doosan units on order with the first permanent pair targeted for early 2027. That trajectory reaches roughly 4–4.3 GW by March 2027, powering on the order of 830,000 deployed chips. Treat the monthly cadence as a projection, not a disclosure — the S-1 gives quantities without dates.
The counterweight is legal, not logistical. The NAACP and environmental groups’ lawsuit was amended in June after discovery found nearly 30 additional gas turbines at the Southaven site — 57 total — with emissions claims under the Clean Air Act. An adverse ruling on turbine operation is the single most direct threat to the power ramp. However, no environmental rulings have required removal of turbines. Usually fines and remediation. SpaceXAI have various environmental waivers and permits from the states and a statement that they are national security critical from the DOJ.
The 2027 proforma
Assuming ~4 GW exit capacity, roughly half leased externally, a $46B/GW base rate on new leases, internal inference revenue growing from a ~$4B run-rate (Grok, Cursor), cash opex near 20% of revenue, and ~$15B of D&A:
Three readings matter. First, even the bear case is operating-profit positive, because the signed $27.8B run-rate alone covers total costs. Second, fiscal-year revenue is much lower than exit run-rate — Q4 2027 annualized revenue in the base case is $90–105B, and conflating the two is the most common error in bullish writeups.
Third, company-level free cash flow is a different question entirely: continuing capex of $50–70B/year keeps FCF negative-to-breakeven except in the bull case. “Proforma cashflow positive on already-built assets” is true and meaningful for return on invested capital; it is not the same as the company generating cash.
Where the risk actually lives
Not in 2027 — in 2028 or later. The two anchor contracts carry 90-day termination rights after their initial periods, and both of those tenants are spending historic sums to stop needing rented compute.
Alphabet raised $84.75B in equity in June explicitly for its own AI infrastructure, on top of $180–190B of 2026 capex. Today’s $50B+/GW rates are scarcity pricing. The bet embedded in every number above the signed line is that scarcity persists through the 2027 signing window. Reflection is the encouraging counterpoint — at $1.8B/year it is only 6% of the signed base, but it is the first tenant that isn’t a hyperscaler building an exit ramp, and the AI-lab tier is where durable second-cycle demand would come from. The turbine litigation is the second-order risk; depreciation lives that prove too generous for fast- obsoleting accelerators are the third.
Verdict
Bullish on 2027 operating economics — the contracted base makes the near-term case a
matter of delivery, not forecasting. Cautious on rate durability past 2027, when the bridge contracts hit their exit windows in what should be a better-supplied market.
The “capex written off over 5–6 years creates positive earnings” thesis is validated for this asset base. The unresolved question is what the second lease cycle prices at.
Sources: SpaceX S-1 and amended S-1 filings, Alphabet SEC filings and investor
presentation (June 2026), CNBC, TechCrunch, DataCenterDynamics, Epoch AI, Memphis
Commercial Appeal (June 11, 2026). Power-ramp projections via NextBigFuture (July 7, 2026.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.