Bernstein Bullish on AI Despite Market Downturn! Raises Server and AI Chip Shipment Forecasts, Calls 2026 the Turning Point for ASICs

Bernstein's latest report indicates that despite a sharp correction in AI supply chain stocks in July, the four major cloud providers continue to raise capital expenditures, driving stronger demand for AI servers and AI chips. As a result, Bernstein has raised its shipment forecasts for 2025 to 2028, projecting the global server market to surpass $1 trillion by 2028. Specialized application-specific integrated circuits (ASICs), high-bandwidth memory (HBM), NVIDIA's (NVDA-US) Rubin, and Google's (GOOGL-US) TPU are expected to be the main growth drivers.

Although AI supply chain stocks in the U.S. market have seen a notable pullback since July, investment firm Bernstein has逆势 raised its shipment forecasts for servers and AI chips in its latest report.

In this report titled '2Q26 AI Server Pulse: Summertime Cool-Down,' Bernstein points out that the recent stock price correction was primarily driven by three factors:

Market rumors suggesting NVIDIA's next-generation Kyber rack may face delays;

Investor concerns over whether overall AI compute infrastructure has become overbuilt;

And momentum funds that had poured into related stocks over the past few months beginning a broad deleveraging.

Analysts describe this correction as a 'summertime cool-down' in the industry's economic cycle, implying it is a short-term emotional pullback rather than a sign of weakening fundamentals.

Major Cloud Providers Continue to Increase Capital Expenditures

The core rationale supporting Bernstein's optimistic shipment forecasts comes from the continued upward revisions in cloud providers' capital expenditure plans.

The report notes that after the Q2 earnings announcements, the market consensus forecast for 2026 capital expenditures by the four U.S. cloud giants—Amazon (AMZN-US), Microsoft (MSFT-US), Google, and Meta (META-US)—has been raised by approximately 15% compared to March, with expectations of a 58% compound annual growth rate from 2025 to 2027, totaling nearly $940 billion.

Including emerging cloud providers such as Oracle (ORCL-US), CoreWeave (CRWV-US), and Nebius (YNDX-US), the combined capital expenditure in 2027 could reach $1.11 trillion.

The same report also shows that the total global investment in under-construction and planned data centers has risen from approximately $960 billion in April to about $1.18 trillion, reflecting that industry confidence in long-term computing demand remains unshaken despite short-term stock price volatility.

Global Server Market to Surpass $1 Trillion by 2028

On this basis, Bernstein has raised its compound annual growth rate forecast for global server shipments from 2025 to 2028 to 15%, with the growth rate for high-end GPU-equipped AI servers raised even higher to 22%.

If this model holds, the global server market size will surpass $560 billion in 2026 and cross the $1 trillion threshold by 2028.

For high-end AI servers based on an eight-GPU configuration, 2026 shipments are forecast to grow 48% year-on-year, with growth moderating to the mid-teens range (13%-17%) in 2027. In 2028, due to NVIDIA's next-generation Rubin Ultra adopting a dual-die design, shipment growth is expected to converge to high single digits.

The report specifically notes that with the rise of 'AI agent' applications, inference workloads have evolved from single-step queries to continuous multi-step processes, driving demand for CPU-centric coordination computing. This will allow general-purpose server shipments to maintain mid-double-digit compound annual growth through 2028.

Rack-Scale Deployment Becomes Mainstream, Kyber Yield Remains a Variable

As AI server shipments rapidly shift from individual boards to full rack-scale deployments, Bernstein estimates NVIDIA's full rack shipments will grow from approximately 61,000 units in 2026 to about 88,000 units in 2027.

The report states that rack-scale deployment will officially become the market mainstream this year, with traditional HGX/MGX board formats expected to consume only about 35% of total GPU volume. In aggregate, the total number of NVIDIA chips actually deployed downstream this year is estimated to reach approximately 6.7 million.

On the product timeline, GB300 racks began volume shipments in Q4 last year, while NVIDIA's next-generation Vera Rubin racks and AMD's (AMD-US) Helios racks are both targeting volume production in Q4 this year. Helios' customer list includes Meta, Microsoft, Oracle, OpenAI, and Anthropic.

However, the report also highlights potential risks. The next-generation Kyber rack's PCB backplane technology for connecting compute blade modules remains a key bottleneck. NVIDIA is expected to require one to two additional quarters of collaboration with major PCB suppliers to finalize design and yield improvements before mass production can proceed smoothly.

Custom Chips Accelerate Rise, 2026 Marks the ASIC Inflection Point

Beyond GPU servers, Bernstein also identifies this year as a pivotal turning point for the ASIC industry.

Cloud providers are shifting from small-scale pilots to large-scale deployment of in-house developed chips to reduce per-unit compute costs, manage total cost of ownership, and decrease reliance on external GPU supply chains.

The report estimates that ASICs will account for about 45% of total AI chip shipments using CoWoS advanced packaging this year, with Google TPU shipments expected to grow over 90% year-on-year, representing about 42% of CoWoS-based ASIC shipments.

In terms of value, ASICs' share of the overall XPU (GPU + ASIC) market is projected to rise from 10% last year to 20% by 2027, despite NVIDIA's continued leadership in the training segment.

The overall XPU market size (including HBM) is estimated to expand from approximately $190 billion last year to about $500 billion by 2027. The standalone data center ASIC market is projected to reach between $140 billion and $160 billion by 2028, with about half of that value attributed to HBM.

In terms of competitive dynamics, Google TPU is expected to capture nearly 70% of the overall ASIC market in 2027 and around 60% in 2028. For TPU alone, the corresponding market size in 2028 is estimated to be at least $90 billion.

Supply chain signals also align:

Broadcom (AVGO-US) has signed a long-term contract with Google to develop and supply multiple future generations of TPU through 2031, while expanding its collaboration with Meta;

Marvell Technology (MRVL-US)'s custom accelerators are now being mass shipped across all four major cloud providers;

MediaTek (2454-TW) has secured two Google TPU projects (v8t and v9 generations) and has raised its ASIC revenue outlook for this year to 'over $2 billion,' while also projecting the 2027 ASIC addressable market to reach $80 billion, with a target market share of 15%-20%.

On chip specifications, Google's Ironwood (i.e., TPU v7) delivers double the performance per watt and six times the HBM capacity compared to the previous generation, with a single rack capable of integrating 256 accelerators; Amazon's Trainium 3 offers FP8 compute performance roughly twice that of the previous Trainium 2, and the next-generation Trainium 4 is expected to deliver BF16 compute power approaching 3900 TFLOPS per chip.

Supply Chain Data Cross-Validates, Valuation and Profitability Diverge

Bernstein also cross-validates the reasonableness of the above shipment forecasts through revenue performance at the supply chain level.

Wiwynn and Accton, deeply involved in Amazon and Meta's ASIC projects, are consensus-forecasted to see revenue growth of 36% and 54% respectively this year; Alchip (3661-TW), which saw a 41% revenue decline last year due to the end of a U.S. IDM project, is expected to see revenue nearly double this year driven by the Trainium 3 project. Monthly revenue growth for Taiwan's ASIC-related companies also re-accelerated in June this year.

Additionally, Taiwan's exports of ICT and audiovisual products encompassing AI servers saw a 63% quarter-on-quarter increase in Q2, primarily driven by simultaneous volume ramp-ups of GB300, ASIC, and general-purpose servers; although ODM/OEM manufacturers' inventory levels have continued to rise since 2024,