As Broadcom Plunges, an Investment Bank Drops a "Bombshell": Channel Checks Show Google TPU Shipments to Reach 35 Million Units by 2028, Far Exceeding Expectations!

Amid pressure on Broadcom's stock price, a circulating channel check report from Mizuho recently sent a strong bullish signal for the company. The core data point is singular but explosive: Forecasting that Google TPU shipments will exceed 35 million units by 2028.

What does this number imply? Google TPU shipments were approximately 2.4 million units in 2025, with an estimated 4.3 million units in 2026, jumping to 35 million units by 2028— an eightfold increase over two years, implying a compound annual growth rate (CAGR) of approximately 185%.

Google's TPUs are custom AI chips (ASICs), and Broadcom is its core design partner. For every additional TPU shipped, Broadcom generates more revenue from its ASIC business. Simply put, Google provides the demand and architecture, while Broadcom handles chip design services and packaging, before handing off manufacturing to TSMC. Broadcom earns corresponding design service revenue for each TPU shipped.

However, the credibility of this figure has been directly questioned by market observers: The total industry forecast for ASIC volume in 2028 is only 15 million units; is Google alone accounting for 2.3 times that amount? "I am genuinely curious whether some intern at Mizuho had Claude extrapolate a trend line, and no one bothered to verify the output."

35 Million Units: How Outrageous Is This Figure?

The figures in Mizuho's channel check have sparked significant controversy within the industry.

Let's compare several benchmarks:

• Counterpoint Research forecasts that by 2028, the combined ASIC shipments from all hyperscale cloud providers, including Google, Amazon, Microsoft, and Meta, will total approximately 15 million units.

• Channel checks by Wolfe Research indicate Google TPU shipments of approximately 7 million units.

• Mizuho's own model had previously forecast ASIC shipments at around 7 million units as well.

Yet, the figure provided in this channel check is 35 million units— 5 times Mizuho's own model and 2.3 times the total industry forecast.

In other words, according to Mizuho's channel check, Google's TPU shipments in 2028 alone would exceed twice the sum of all ASIC projects globally.

Skepticism: Mathematically Consistent Does Not Mean Physically Feasible

Market observer @LinQingV directly called out and questioned this research on the X platform.

"Have you seriously considered what shipping 35 million TPUs in 2028 implies?" he wrote.

His logic for skepticism is clear: Growing from 4.3 million to 35 million represents a two-year CAGR of about 185%. "Even accounting for MediaTek as a dual supplier and Anthropic's 3.5GW agreement, no global supply chain can expand at this speed."

He also noted that Mizuho's calculation of Micron's HBM demand is internally consistent in terms of numbers: Cumulative TPU shipments from 2026 to 2028: 4.3 million + 10.7 million + 35 million = 50 million units. The math is correct. However, his conclusion is that "internal consistency" and "physical feasibility" are two different things.

He concluded by writing: "I am genuinely curious whether some intern at Mizuho had Claude extrapolate a trend line, and no one bothered to verify the output."

What Else Did Mizuho Mention Besides Google TPUs?

This channel check report does not solely concern Google.

Mizuho also mentioned that OpenAI's ASIC chips are under development, related to Broadcom's "Nexus" project. This project, with a scale of 10GW, will complement the roadmaps for Google TPUs, Meta MTIA, and ARM chips.

Regarding ARM, calculations show that its AI ASIC chip development is progressing, with an expected launch between late 2026 and early 2027.

The impact on Micron was also highlighted separately: If cumulative TPU shipments reach 50 million units between 2026 and 2028, ASICs will occupy an increasingly large share of the HBM market, especially as next-generation ASICs transition to the HBM4e specification.