Morgan Stanley: AI Computing Power Enters an Era of GPU and XPU Running in Parallel
Miles Bennett
Morgan Stanley's latest semiconductor report projects the global AI chip market will grow from $485 billion in 2026 to $753 billion by 2030, accounting for roughly half of all chip sales worldwide. This means → the AI race is shifting from 'who has the best GPU' to 'who can build the most complete chip ecosystem.'
GPUs alone aren't enough — what are XPUs?
AI's center of gravity is moving from training models to inference — actually running them to answer questions and make decisions. GPUs remain essential for training, but inference workloads are far more varied and call for specialized chips.
XPUs — a catch-all for AI-specific processors including AI ASICs, NPUs, and others — are rising fast. This means → cloud providers are no longer just buying GPUs; they need large fleets of purpose-built chips to cut inference costs and boost efficiency.
In plain terms = a GPU is a Swiss Army knife — it does everything but isn't cheap. XPUs are a set of dedicated tools, each doing one job faster and cheaper. Running both in parallel is becoming the standard setup.
Where is the cloud money going?
Amazon, Google, Microsoft, and Meta boosted Q1 2026 capex by 95% year-over-year; capex as a share of EBITDA held at roughly 50%.
Under Morgan Stanley's supply-chain-driven bull case, 2026 cloud capex reaches $796 billion, of which AI servers account for about $600 billion and cloud AI ASICs plus non-Nvidia GPUs roughly $90 billion.
This reflects a strategic bet: cloud giants now treat AI infrastructure as their top long-term priority — not just buying chips, but building proprietary computing stacks.
Why do packaging and manufacturing suddenly matter so much?
The performance bottleneck is shifting from chip design to how chips are assembled together. TSMC's CoWoS — an advanced packaging technology that bonds multiple chips tightly onto one substrate — will keep expanding capacity into 2027, and SoIC, TSMC's chip-stacking technology, is a key growth area ahead.
This means → the competitive battleground is no longer the chip alone but the full chain: wafer fabrication, advanced packaging, testing, and system integration.
Morgan Stanley flags a risk, however: rising wafer, OSAT (outsourced assembly and test), and memory costs — plus AI crowding out resources for non-AI chips — could squeeze chip-design companies' margins in 2026.
How high can China's AI-chip self-sufficiency go?
Morgan Stanley estimates China's AI GPU market will reach roughly $91 billion by 2030, with domestic AI-chip self-sufficiency rising to about 70%.
DeepSeek proved that low-cost AI inference is viable, accelerating inference demand and opening space for China's domestic AI chip supply chain.
This means → as China's advanced-node capacity expands, homegrown chips will grow increasingly competitive in inference workloads, and AI infrastructure buildout will lean more heavily on local supply chains.
Who really wins in the AI era?
Morgan Stanley's core call: the AI era's winners won't be GPU makers alone but players spanning chip design, advanced manufacturing, advanced packaging, testing, and AI-specific processors.
Put simply = the old question was "whose GPU is fastest"; the new question is "whose entire computing ecosystem is most complete."
This signals that AI semiconductor competition has entered a systems-level phase — no single technology path is decisive anymore. Ecosystem completeness is the variable that matters most.
Content is for reference only, not financial advice.