Marvell's $12.2B Warrants from Google: Filling TPU Ecosystem Gaps, Not Replacing Broadcom
Nashnova编辑部
Google granted Marvell a $12.2 billion warrant covering five custom chip categories, but the 8-K's key word is "attach" to the TPU ecosystem — this means Marvell is filling a gap, not taking Broadcom's work.
What does this deal actually cover?
Marvell's 8-K filing on August 19 lists five custom semiconductor product categories, with the critical phrasing being "attach" to the TPU ecosystem.
This means → Marvell wraps around the TPU as a companion supplier. It does not enter TPU chip design itself.
DIGITIMES senior analyst Luke Lin stated plainly: "The scope described is around the TPU ecosystem, not inside it."
Which four categories are "undefended"?
The first four — storage controllers, network interface controllers, memory interface controllers, and near-memory compute chips — are companion silicon that determines whether a TPU rack can scale.
In plain terms = these are not the "brain" but the "nerves and blood vessels": they decide how fast data reaches the TPU and how fast results get out.
Marvell has deep expertise in high-speed I/O and interconnect. Its management argued at COMPUTEX 2026 that "interconnect, not compute, will be the next bottleneck in AI data centers." These four categories lacked a strong competitor in Google's supply chain.
What about the fifth category — the inference accelerator?
The fifth product is an AI inference accelerator — a compute chip, not companion silicon.
Reports suggest the project may include Google's "Merope" language processing unit (LPU) inference chip. KeyBanc estimates the project could generate $20–25 billion in revenue during 2028–2029, but neither company has confirmed the figure.
This means → Marvell has won an entirely new compute project, not a transfer of existing business from Broadcom or any other supplier.
What do the warrant terms reveal about bargaining power?
Google can buy up to roughly 58.97 million shares of Marvell common stock at $206.58 per share, near the pre-announcement market price. Total exercise value: approximately $12.18 billion, or about 6.7% of shares outstanding.
Roughly 1.36 million shares vest on a time schedule. The remaining ~57.61 million shares vest on a purchase-linked basis — one tranche per $500 million in custom product revenue, 240 tranches in total, running from fiscal Q3 2027 through fiscal year-end 2033. The filing discloses no minimum purchase commitment.
For comparison, AMD previously granted OpenAI and Meta 160 million shares each at just one cent per share — essentially free. In plain terms = AMD needed flagship-customer endorsement and paid heavily in equity; Marvell gave up far less, signaling it did not need to "buy" this business with stock.
Did Broadcom actually "lose the contract"?
After the announcement, Broadcom fell 3% while Marvell rose 13% to $243.66. The market narrative quickly became "Broadcom lost the deal."
But the 8-K points to supplier diversification, not a confirmed loss of Broadcom shipment volume. Broadcom is designing the "Sunfish" TPU v8 training chip; MediaTek is handling the "Zebrafish" low-cost inference variant. Both target TSMC 2 nm with mass production in late 2027. These projects are unaffected.
This reflects a simpler reality: Google's TPU-related silicon purchases are growing fast enough to support a second design house — not to replace an existing one. Broadcom's agreement with Google runs through 2031.
Who is the most certain beneficiary?
Marvell is fabless. TSMC is the most likely manufacturer for these designs — just as it is for Sunfish and Zebrafish.
This means → regardless of how Google's AI chip supply chain expands or how the pie is split among designers, TSMC collects on every chip at the manufacturing end.
The filing contains no indication that MediaTek competed for the business Marvell won.
Content is for reference only, not financial advice.