SemiAnalysis Reframes H2: Microsoft's 10GW Awakening and a $300B ARR Path for SpaceX

Microsoft has signed 10GW of binding data center contracts year to date, about $300B in new commitments. SpaceX is modeled to reach 10GW by end 2027, a path to $300B ARR at $100M per MW per year. The buy side will argue about this for weeks.

Microsoft just woke up. SpaceX is trying to become a compute company in eight quarters. SemiAnalysis tied the two together on Aug 7 in SpaceX 10GW in 2027, and the framework is now the H2 debate for the whole AI supply chain.

The core claims are simple and large. Microsoft has signed 10GW of binding data center contracts year to date in 2026, roughly $300B in new commitments, reversing its 2024 to 2025 leasing pause. The economic anchor is about $100M per MW per year of inference revenue, or $100B per GW per year. SpaceX is modeled as feasible to reach 10GW plus by end of 2027, a path to $300B ARR even if only 50 percent of incremental compute is monetized, with Microsoft as its largest offtaker.

Source: SemiAnalysis Datacenter Model via SpaceX 10GW in 2027, Aug 7 2026. 2024 to 2025 bar illustrative of pause period, 2026 YTD is 10GW binding.

1. Microsoft's 10GW awakening

SemiAnalysis describes 2026 as Microsoft's awakening after a deliberate pause in leasing through 2024 and 2025. Year to date, Microsoft has signed 10GW of binding contracts across four buckets: behind the meter power agreements, neocloud offtakes, third party preleasing, and self build.

The dollar tag is about $300B in new commitments year to date. That is not a pipeline or a letter of intent. SemiAnalysis frames these as binding contracts, which is why the note landed so hard.

For context, MSFT closed at $506.06 on Aug 10 and traded at $502.20 on Aug 11, down 0.76 percent intraday, after a 31.4 percent run over the prior 21 trading days. The stock is pricing that Microsoft is no longer capacity constrained on purpose.

Source: Bargo equity_price_history, closes through Aug 10. MSFT spiked 15.5 percent on Jul 30 after earnings, then added another 12 percent into Aug 10.

2. The $100M per MW per year anchor

The number that makes the rest of the math work is inference revenue per unit of power.

SemiAnalysis anchors inference at about $100M per MW per year, or $100B per GW per year, on GB300 NVL72 running agentic traces at about 58 percent interactivity and under 6 seconds time to first token. The note shows GB200 at $73.4M per MW per year and GB300 at $99.7M per MW per year, so the $100M figure is the GB300 case.

That revenue climbs sharply up the stack. SemiAnalysis shows revenue per GW rising about 8 times from neocloud 5 year IaaS rental at the bottom to model layer tokens at the top, with SpaceX pricing between the two.

Source: SemiAnalysis Tokenomics Model and Inference Simulator via SpaceX 10GW in 2027. Multiples shown vs neocloud 5 year IaaS average.

At $100M per MW per year, the Microsoft math is clean. 10GW contracted equals $1T of potential inference revenue capacity if fully monetized at that rate. SemiAnalysis does not assume full monetization. The illustrative case that drives the Azure debate is smaller: if Microsoft contracts 3GW from SpaceX, fully live by year end 2027 at $100M per MW per year, that is $300B of exit ARR in calendar 2027.

SemiAnalysis translates that into Azure growth. On a $217B base forecast, plus $150B of GPT token revenue, Azure year over year growth peaks at 179 percent in 4Q27 in the illustrative scenario. The table in the note shows 1Q26 at 40 percent year over year, then triple digit prints through 2027.

Source: SemiAnalysis Tokenomics Model, illustrative scenario where Microsoft contracts 3GW from SpaceX at $100M per MW per year. 4Q27 peak is 179 percent.

3. SpaceX 10GW in 2027 and the $300B ARR path

The SpaceX side is more aggressive. SemiAnalysis models SpaceX conservatively building and delivering an incremental 6 to 8GW in 2027 alone, with potential to exceed 10GW. At an estimated $50B per GW of capex, that is $300B to $500B of capex in 2027, putting SpaceX on par with AWS and Google.

The ARR path follows directly from the $100M per MW anchor, but with a haircut. SemiAnalysis assumes only 50 percent of 2027 incremental compute is monetized, the remainder reserved for Grok and Cursor teams for training with no inference revenue modeled. Even with that haircut, the path is $300B ARR by end of 2027.

Shared by @SemiAnalysis_ and @pequityresearch, the summary table is: 6 to 8GW incremental in 2027, 3 to 4GW monetized at $100M per MW per year, equals $300B to $400B ARR, with 10GW total pushing toward $500B if fully monetized.

Source: SemiAnalysis via SpaceX 10GW in 2027 and pequityresearch summary. Assumes 50 percent of incremental GW is monetized at $100M per MW per year.

Financing is expected via Nvidia vendor financing, which SemiAnalysis links to Elon Musk's focus on being Nvidia exclusive, plus industry high pricing and rapid deployment timelines of 3 to 5 months lead time. The note also flags a 90 day cancellation provision that lets SpaceX reclaim compute if Grok needs spike, which matters for how binding the Microsoft offtake really is.

SpaceX turning into a compute company shows up in the revenue mix. SemiAnalysis shows AI compute going from nothing to 77 percent of run rate revenue on an annualized exit month times 12 basis, with 2026 anchored to signed contracts and 2027 illustrative at the 10GW path.

4. Why the buy side will argue about this for weeks

This framework is bullish the whole supply chain if you believe the $100M per MW per year holds, and it is a capex warning if you do not.

If Microsoft has truly signed 10GW binding, the shortage is not easing. It is spreading from GPUs into power, data center shells, and memory. That is consistent with the $90B of new chip capex approved or priced in the same week, from TSMC and SK Hynix funding expansion from cash flow to Intel printing shares to fund foundry, covered in The Supply Side Voted With Its Wallet.

For Nvidia, the read through is direct. NVDA closed at $217.55 on Aug 10 and traded at $220.76 on Aug 11, up 1.48 percent intraday, with a $5.27T market cap and forward P/E of 16.9. Every GW at $50B of capex is Nvidia content, and every GW at $100M per MW per year is inference demand that needs that content. The bear case is that $100M per MW per year assumes value based pricing and high utilization that may not hold if token prices fall faster than volume grows.

For Microsoft, the debate is whether Azure can actually grow triple digits without breaking gross margin. SemiAnalysis models it as feasible because inference gross margin expands with GB300, but the market will want to see it in reported Azure growth and capex guidance.

For SpaceX, the debate is execution and funding. 10GW in 12 months is more than 2 times AWS global delivery pace in 2025, concentrated in one company, and it needs roughly 9 to 10GW of gross site power and sixteen 500MW modules built concurrently. As one critique noted, announcements are not energized capacity, and Colossus I mixed extraordinary xAI execution with real utility help that may not scale.

5. What to watch

• Microsoft earnings and capex guide. Does management confirm the 10GW binding figure and raise capex to match, or does it stay vague on power contracted versus power energized.
• SpaceX delivery pace. Watch for site level power agreements and energization dates, not just shell announcements. 3 to 5 month lead times only matter if power is live.
• The $100M per MW per year print. GB300 pricing at $99.7M per MW per year is the anchor. If realized pricing slips toward GB200 levels at $73.4M, the $300B ARR path needs more GW to hold.
• Nvidia vendor financing terms. If SpaceX funding is tied to Nvidia exclusivity and high pricing, watch for disclosure on payment terms and cancellation provisions.
• Memory and power as bottlenecks. If HBM and enterprise SSD stay tight and power remains the constraint, the infrastructure financing model for memory and data centers pays back faster.

Sources

• SemiAnalysis, SpaceX 10GW in 2027: Why It Is Real, Will Drive $300B ARR for SpaceX, and Why Microsoft Will Be the Largest Offtaker, Aug 7 2026 SemiAnalysis
• SemiAnalysis on X SemiAnalysis_
• pequityresearch summary thread pequityresearch
• Bargo market data: MSFT $506.06 close Aug 10, $502.20 intraday Aug 11, NVDA $217.55 close Aug 10, $220.76 intraday Aug 11, via Alpaca SIP