Google Reportedly Partners with AMD to Design 10th-Gen TPU; Wedbush Says It Could Reshape the Chip Industry Landscape
Nashnova编辑部
SemiAnalysis reports that Google is working with AMD on a 10th-generation TPU — its custom AI chip. If confirmed, the move would mark Google's first use of an outside CPU vendor in core chip design, a shift Wedbush calls 'seismic' for the semiconductor industry.
What are Google and AMD actually building?
Chip-industry research firm SemiAnalysis reports that Google is partnering with AMD to design the 10th-generation TPU — a tensor processing unit, Google's in-house AI chip.
Google is tapping AMD's expertise in CPU architecture, advanced packaging, and network interconnects.
This means → Google is no longer going fully in-house on AI silicon. It is pulling an external CPU maker into the core design loop.
Neither Google nor AMD has commented. The report remains unconfirmed.
Why does Wall Street call this a "seismic shift"?
Wedbush Securities analyst Matt Bryson wrote in a client note that the market had expected Google to partner with Broadcom, Marvell, or Intel — AMD was not in the consensus picture.
In plain terms = Google skipped the three front-runners and picked a name few were betting on. That is why Bryson calls it seismic.
This reflects a broader shakeup: incumbents in custom AI silicon can no longer take their positions for granted.
What two industry trends does this reveal?
Bryson argues that even as a rumor, the report already surfaces two structural trends:
Trend one: the strategic value of ASIC — application-specific integrated circuit — design capability keeps rising. This means → whoever can build custom chips for hyperscalers holds the pricing power.
Trend two: IP blocks — reusable chip-design modules — are becoming critical. In plain terms = chip design increasingly works like building with standardized Lego bricks. The vendor with the most bricks ships the fastest.
Where does Google's in-house chip program stand today?
Google launched its 8th-gen TPU in April this year — TPU 8t for model training and TPU 8i for inference.
A 10th-gen chip co-designed with AMD would skip at least one full generation, suggesting the project is still in early-stage R&D.
This means → any material impact on AMD's revenue or Google's AI infrastructure roadmap cannot be assessed until both companies confirm the partnership. It is too early to price in earnings.
Is the custom-chip race a zero-sum game?
Bryson is explicit: under today's constrained compute supply, custom-chip competition is not zero-sum.
In plain terms = demand for AI chips far outstrips supply. AMD winning a deal does not mean Broadcom loses one — the pie itself is still growing.
This reflects the depth of today's AI compute scarcity: demand growth is outpacing capacity expansion, and multiple vendors can benefit at the same time.
Content is for reference only, not financial advice.