Morgan Stanley sees Google, Amazon, Meta, Microsoft adding 34GW of AI compute by 2027. The forecast boosts Nvidia and custom chip demand but raises execution risk.
Morgan Stanley analysts project that Morgan Stanley estimates that Google, Amazon, Meta Platforms, and Microsoft could collectively add up to 34 gigawatts of compute capacity by 2027. The forecast** reported by Seeking Alpha targets the AI infrastructure buildout that has become the central capex theme for the largest cloud and internet companies. If realized, 34GW would represent a massive step-up from current levels, with direct implications for chip suppliers, power infrastructure, and the pace of AI model deployment.
The scale of the number matters. 34GW is roughly the output of 30 large nuclear reactors running continuously. That level of compute expansion suggests hyperscalers are planning for an order-of-magnitude increase in AI training and inference workloads. The Morgan Stanley projection reinforces the thesis that the current capex cycle is not a one-quarter spike but a multiyear commitment. For traders watching the Nvidia (NVDA) ecosystem and the broader semiconductor supply chain, the 34GW figure provides a concrete planning assumption to stress-test against.
The analyst note reportedly cites demand for Nvidia GPUs and custom chips, the two key levers hyperscalers use to scale compute. Nvidia remains the primary beneficiary for general-purpose AI training clusters. Companies like Google (TPU), Amazon (Trainium, Inferentia), and Meta (MTIA) are increasingly designing their own ASICs to optimize cost and power per inference. A 34GW capacity target implies that both GPU procurement and custom chip tape-outs will need to ramp sharply through 2026 and 2027.
Broadcom and Marvell Technology are the primary merchant silicon partners for these custom chip programs. The 34GW forecast does not specify how the split between GPUs and custom silicon will evolve. Yet the sheer scale of the buildout suggests that both tracks will run in parallel. Nvidia’s next-generation Blackwell architecture and the custom chip roadmaps from each hyperscaler will compete for fab capacity at TSMC and Samsung. Any supply constraint at the foundry level would become a bottleneck for the 34GW target.
The bullish read on this forecast is straightforward: more compute means more GPU shipments, more networking gear, and more data center construction. The skeptical read, which AlphaScala’s framework demands, focuses on execution risk and potential overbuild. Capex for the four companies in aggregate is already expected to exceed $200 billion in 2025. Adding 34GW of capacity by 2027 would push that number higher. The question is whether end demand – from enterprise AI adoption, consumer AI features, and autonomous systems – will fill the compute before depreciation costs eat into margins.
Meta provides a cautionary case. The company increased its 2024 capex guidance multiple times as it acquired Nvidia H100 GPUs. Its Alpha Score of 54/100 (Mixed) suggests the market prices in execution risk even as the growth story remains intact. Microsoft, with an Alpha Score of 50/100 (Mixed), shows a similar tension: its Azure AI business is growing rapidly, yet the capital intensity of building out that capacity is visible in free cash flow compression. The 34GW forecast does not resolve that tension; it amplifies it.
Traders evaluating the hyperscaler capex theme should watch three signals. First, Q2 2025 earnings commentary from the four companies: any reduction in capacity planning guidance would break the 34GW assumption. Second, custom chip development milestones – tape-outs for Google’s next TPU or Amazon’s Trainium 3 – confirm the shift toward in-house silicon. Third, power availability: data center interconnection queues in Virginia, Texas, and the Pacific Northwest show already-strained utility grids. A 34GW expansion requires utilities to permit, build, and deliver that power on time, a process that historically takes five to seven years.
Morgan Stanley’s 34GW figure is a useful anchor for modeling the AI compute cycle. It is not an investment thesis on its own. The next decision point comes when hyperscalers report second-quarter results and update their full-year capex outlooks. Any deviation from the projected ramp will ripple through Nvidia, the custom chip names, and the utilities serving data center regions.
AlphaScala provides stock scores and market analysis for the MSFT stock page and META stock page, along with broader market analysis coverage.
Prepared with AlphaScala research tooling and grounded in primary market data: live prices, fundamentals, SEC filings, hedge-fund holdings, and insider activity. Each story is checked against AlphaScala publishing rules before release. Educational coverage, not personalized advice.