Goldman Sachs Raises AI Server Demand Forecast, Projects $1.1 Trillion Market by 2028
Miles Bennett
Goldman Sachs on June 24 sharply raised its global server market forecast, projecting total 2028 revenue of $1.1 trillion with AI server rack revenue growing at a 118% CAGR to $561.4 billion — rising memory costs and ballooning AI workloads are pushing the entire supply chain's pricing center upward.
Where does the money go inside a $1.1 trillion market?
Goldman projects 2028 global server revenue at $1.1 trillion. AI server racks alone account for 51% of that — $561.4 billion.
This means → in three years, more than half of every dollar spent on servers worldwide goes to AI. AI is no longer the growth add-on; it is the main event.
Two forces drive the upgrade: high memory costs and relentless AI workload growth. In plain terms = models keep getting larger, demanding more memory, while memory itself is getting pricier — the two compound, lifting the market's revenue ceiling.
Why are custom ASIC chips gaining on GPUs?
Goldman expects ASIC — application-specific integrated circuits, chips custom-built for a single task — to rise from 50% of total AI chip volume in 2026 to 55% in 2028.
In unit terms: ASIC demand jumps from 9.34 million units in 2026 to 17.65 million in 2028; GPU demand grows from 9.51 million to 14.57 million — ASIC is scaling faster.
This means → cloud providers are shifting more budget from general-purpose GPUs to custom ASICs. The logic is straightforward: ASICs deliver lower latency and cut expensive memory read-write cycles. Saving memory is saving money.
How much did Nvidia rack shipments get raised?
Goldman lifted its NVL72-equivalent rack shipment forecast to 55,000 / 105,000 / 163,000 units for 2026–2028. The 2027 and 2028 figures are 16% and 20% above prior estimates.
Nvidia NVL72 racks specifically are now projected at 92,000 and 148,000 units in 2027–2028, roughly 19–22% above the old forecast.
AMD rack shipments are essentially unchanged at 5,000 / 13,000 / 15,000. This reflects that in the high-end AI rack market, Nvidia remains the dominant source of incremental volume; AMD has not yet captured additional share.
Why is Foxconn set to command nearly 70% of ODM share?
Goldman projects Foxconn / FII's share of the AI server rack ODM — original design manufacturing, where the contract manufacturer designs and ships directly — market will rise from 55% in 2026 to 69% in 2028.
In plain terms = for every 10 AI server racks built globally by 2028, nearly 7 come from Foxconn. The ODM-direct model — factory ships straight to the cloud provider — remains the dominant supply path.
This means → manufacturing in the AI server chain is concentrating fast. Foxconn's scale advantage compounds in rack assembly, a capital-heavy link where size matters most.
US vs. China cloud capex — who is ramping harder?
US hyperscalers (Microsoft, Amazon, Meta, Google, Oracle) are projected to raise capex by 76% / 35% / 8% YoY from 2026 to 2028. Three-year forecasts were revised up 8% / 27% / 25%, bringing the 2028 cumulative total to $1.145 trillion.
Chinese hyperscalers (ByteDance, Alibaba, Tencent, Baidu) see YoY growth of 80% / 20% / 18% over the same period, with forecasts revised up 37% / 44% / 55% — markedly steeper upward revisions than their US peers.
This means → Chinese cloud providers are ramping faster from a catch-up position. Goldman notes that China's chip mix is more dispersed across local suppliers, with GPU and ASIC configurations spread more broadly — unlike the US, where spending concentrates heavily on Nvidia.
How long can this AI infrastructure investment cycle last?
Goldman maintains its bullish call: the AI infrastructure capex cycle extends at least through 2028, with buy ratings on multiple names across the supply chain.
Chips (GPU, ASIC, CPU) remain the largest spending category, followed by HBM — high-bandwidth memory — then networking and other components.
Goldman flags one key variable: whether ASIC penetration continues rising as projected. In plain terms = if cloud providers ultimately do not shift to ASICs at the expected scale, the entire upward-revision logic chain needs to be recalculated.
Content is for reference only, not financial advice.