Google and Marvell Customize AI Chips with Equity Links: How One Agreement Reshapes Cloud Vendors' Supply Chain Choices
CoinMeta
3h ago
Ai Focus
According to the 8-K filing submitted to the U.S. Securities and Exchange Commission by Marvell, the company has expanded its custom semiconductor collaboration with Google and granted Google up to 58,970,907 warrants at an exercise price of $206.58 per share. The agreement covers a range of custom chips connected to the TPU ecosystem, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory computing products. While the market may easily focus on the prominent figure of "up to approximately $12.2 billion in equity value," what is truly important about this arrangement is that most of these rights are not obtained immediately and unconditionally, but are tied to actual purchase revenue over the coming years.
Helpful
No.Help

Marvell向美国证券交易委员会提交的8-K文件披露,公司与Google扩大定制半导体合作,并向Google授予最多5897.0907万股的认股权证,执行价为每股206.58美元。协议覆盖与TPU生态连接的多类定制芯片,包括AI推理加速器、存储控制器、网络接口控制器、内存接口控制器和近内存计算产品。市场容易只看到“最高约122亿美元股权价值”的醒目数字,但这份安排真正重要的,是大部分权利并非立即无条件获得,而是与未来多年实际采购收入挂钩。

文件显示,双方在7月29日签订商业协议,Marvell于8月18日发行认股权证。约136.09万股属于时间归属部分,在首年按季度等额归属;其余股份按照Google及关联方从Marvell 2027财年第三季度至2033财年末的自主采购逐步归属,共分240个等额批次,每实现5亿美元定制产品收入归属一批。认股权证在符合条款和归属条件后可行使,期限到2033年8月18日。换言之, headline中的巨大金额是“最多可买多少股乘以执行价”的名义规模,并不代表Google已经投资相同金额,也不代表Marvell已经锁定全部订单。

从单颗加速器转向整套数据通路

AI基础设施竞争正从“谁有更快的计算核心”转向整机架和数据中心系统效率。推理工作负载的瓶颈可能出现在内存带宽、模型参数搬运、网络拥塞、存储读取或不同芯片之间的数据交换。Google与Marvell披露的合作范围横跨加速、网络、存储和近内存计算,说明超大规模云厂商正在把这些环节作为一个系统共同优化,而不只是采购一颗通用处理器。

定制芯片的吸引力在于可以围绕确定的工作负载削减不必要功能,控制功耗、延迟和单位推理成本。对拥有巨大内部需求的云厂商,哪怕每次调用只节省很小成本,乘以数十亿次请求也会产生可观价值。与此同时,定制方案的代价是研发周期长、流片费用高、需求预测困难,而且一旦架构选择错误,修正速度往往慢于购买成熟通用芯片。因此,Google没有把合作限定在单一器件,而是把多类“连接TPU生态”的产品纳入协议,增加了按项目选择和调整的空间。

认股权证把供应商激励与客户采购绑定。Marvell若想让大部分股份归属,必须形成巨额、持续且合格的定制产品收入;Google则在商业合作成功时获得潜在股权收益。这种结构既能鼓励供应商投入工程资源,也让客户不必在协议签署时一次性承担全部经济承诺。投资者判断时应把“最大潜在股份”“实际已归属股份”“实际采购收入”和“最终行权金额”分开,避免把上限当成确定结果。

对AI芯片产业链意味着什么

第一,超大规模客户正在主动分散设计与供应能力。定制AI芯片不是赢家通吃市场,同一家云厂商可以同时使用内部团队、不同ASIC合作伙伴和通用GPU。新合作不必然意味着原有供应关系立即结束,更可能是为不同代际、不同功能和不同风险等级配置多条路线。对供应商而言,能否提供计算、互连、存储和封装协同能力,将比单个IP模块的峰值指标更重要。

第二,推理正在成为定制化的主战场。训练集群追求极高性能和灵活性,工作负载变化较快;大规模推理一旦稳定,单位成本、能效和服务质量更容易通过专用设计持续优化。协议明确提到AI推理加速器和近内存计算,反映业界在解决“算力增长快于数据搬运效率”的结构性问题。未来评估云厂商资本开支时,不能只数GPU数量,还要观察自研或定制芯片在推理流量中的占比。

第三,供应链财务关系会更复杂。以采购里程碑换取股权上行,能够加深合作,但也带来潜在稀释、客户集中和收入确认风险。Marvell披露的剩余权利按240个批次归属,为市场提供了清晰框架;实际归属仍取决于采购决定。分析公司业绩时,应关注合格收入开始出现的时间、毛利率、研发投入、客户占比及股份稀释,而不是只根据协议理论上限推演估值。

对整个AI产业,这份协议传递的信号比短期股价更持久:云厂商希望掌握从计算到数据移动的系统设计权,同时用长期经济利益确保关键合作伙伴投入。它不会立刻改变现有硬件格局,也不能保证所有定制项目成功;但它证明,下一阶段的竞争单位正在从“芯片”扩大为“工作负载、芯片、互连、内存和资本安排”的组合。谁能把这些要素一起优化,谁才更可能在推理成本竞争中取得优势。

来源:Marvell提交的Form 8-K(2026年8月19日),https://www.sec.gov/Archives/edgar/data/1835632/000119312526356217/d412696d8k.htm

Tip
$0
Like
0
Save
0
Views 13
CoinMeta reminds readers to view blockchain rationally, stay aware of risks, and beware of virtual token issuance and speculation. All content on this site represents market information or related viewpoints only and does not constitute any form of investment advice. If you find sensitive content, please click“Report”,and we will handle it promptly。
Submit
Comment 0
Hot
Latest
No comments yet. Be the first!
Related
CFTC Innovation Advisory Committee's First Meeting: Crypto Assets and Predictive Market Rules Enter the "Technical Details" Phase
The U.S. Commodity Futures Trading Commission (CFTC) held its first meeting of the Innovation Advisory Committee on August 20 in Washington. The committee was restructured from the former Technical Advisory Committee and is tasked with providing recommendations to regulatory authorities on technical, legal, policy, and financial cross-cutting issues. Its members come from exchanges, clearinghouses, crypto companies, predictive markets, academia, and public interest organizations. The meeting itself does not automatically result in new regulations, nor does it equate to consensus among committee members, but it indicates that discussions on derivatives regulation in the United States are shifting from principle-based support for innovation to more actionable issues such as market structure, product boundaries, monitoring capabilities, and customer protection.
币界网
·2026-08-21 10:46:36
63
ArbOS 61 Elara Launched Arbitrum One and Nova: The compliance filtering function is available, but it has not been enabled on the mainnet.
At 17:00 on August 20th ( UTC ), Arbitrum activated ArbOS 61 for Arbitrum One and Nova. This upgrade includes an expansion of the contract capacity for Stylus, improvements to the fee and data availability interfaces, as well as protocol features available for selection by other Arbitrum chains. The most misleading aspect is "protocol-level compliance control": while Elara does incorporate the ability to restrict specific addresses into the software, the official proposal for ArbitrumDAO clearly states that this feature will remain disabled on One and Nova. This upgrade should not be described as the beginning of a general freeze or filtering on the Arbitrum mainnet.
币界网
·2026-08-21 10:46:08
62
In July, Japan's CPI rose to 1.8% and a new base period for 2025 was established: To understand inflation, one cannot focus solely on one figure
On August 21, the Statistics Bureau of Japan's Ministry of Internal Affairs and Communications released consumer price data for July. Nationwide, excluding fresh food, the core CPI rose by 1.8% year-on-year, in line with the median expectations of market surveys, and continued to rebound from 1.6% in June. There is also an important change that is easily overlooked in this release: Japan has officially started using the CPI series with 2025 as the base period. The update of the base period will adjust the weights of goods and services and reset the index reference point to the new consumption structure. Therefore, the July data is not only an inflation reading but also a new starting point for observing changes in residents' spending patterns after the pandemic and the energy shock.
币百科
·2026-08-21 10:45:29
17
Claude Academy Turns AI Training from Prompt Words Class to Judgment Training: What Enterprises Really Lack Is Not "The Ability to Ask", but "The Ability to Verify"
On August 20th, Anthropic officially launched Claude Academy and simultaneously made public its educational approach for AI. On the surface, this appears to be a set of free courses and tutorials; however, what is more noteworthy is the deliberate shift in training focus from "memorizing cue words and techniques" to task judgment, risk grading, result verification, and the division of labor between humans and machines. For companies that are advancing the implementation of large models, this direction is more practical than another round of tool demonstrations: model capabilities change too rapidly, and specific buttons and sentence patterns will become outdated. What can be reused over the long term is rather the employees' ability to decide when to invoke AI, how much context to provide to AI, and at what intensity to review the output.
CoinMeta
·2026-08-21 10:44:42
18
CFTC Discusses Predicting Market Risks and Regulatory Approaches
CFTC The first meeting of the advisory committee focused on predicting market regulation, discussing self-certification, manipulation risks, and issues related to state and federal authorities.
CNBC
·2026-08-21 09:00:59
25
View More