On August 28, Reuters disclosed an "unfinished deal" that brought Anthropic and an AI chip startup named MatX into the spotlight—this company, which competes with OpenAI through its Claude series of large models, seriously considered spending approximately $7 billion to acquire MatX directly, betting to fill its shortcomings in computing power. However, the acquisition negotiations ultimately stopped short of signing, and the specific reasons for the breakdown have not been publicly explained, thus changing the trajectory of the story: rather than placing all its chips on a single target, Anthropic began intensive talks with several chip startups, shifting towards more flexible custom chip collaborations and self-developed routes, hoping to reduce its reliance on NVIDIA GPUs during the high-performance training phase, and to secure a more controllable hardware future for future iterations of Claude. For MatX, the lost acquisition became another test of capital—it is seeking about $4 billion in a new round of financing in the primary market, attempting to capitalize on the "de-NVIDIA-ization" current in a field where Google uses TPUs and Amazon uses Trainium, and where tens of billions of dollars are often needed for investment, positioning itself at the next round of AI chip competition.
$7 Billion Deal Stalled: Acquisition Turns to Withdrawal
Amid increasing pressures on computing power, Anthropic attempted to secure a deal substantial enough to "rewrite the landscape" by integrating chip capabilities into its system. According to Reuters, citing anonymous sources, Anthropic had considered acquiring the AI chip startup MatX for about $7 billion to strengthen its control over underlying hardware at some undisclosed point in time. This figure sharply contrasts with MatX's subsequent funding target of approximately $4 billion in the primary market; if completed, it would instantly pull a startup still at the fundraising table into the hardware sequence of major large model players. This news was disclosed by Reuters between August 27-28 and was subsequently picked up and spread by Chinese tech and crypto media outlets such as Jinse Finance and Lydong, bringing the originally closed negotiations into broader market view.
However, this deal, seen as Anthropic's "purchase of the chip future," never reached the moment of signing. Reuters' report confirmed that the acquisition plan has been abandoned and the negotiations terminated, but the reasons have not been disclosed through any public channels, nor was a complete timeline given, and it was not indicated whether the two parties would maintain other forms of cooperation outside of the acquisition. In the absence of information, the outside world can only see the outcome itself: Anthropic no longer attempts to solve hardware constraints by integrating MatX at once but begins or continues talks with several chip startups within the same time window, exploring self-development or custom chip collaboration paths. Thus, the stalling of the $7 billion deal is not merely a failed acquisition case but marks a crucial turning point in Anthropic's hardware strategy, shifting from "betting on a single company" to "multi-point layout," and it also becomes a watershed for MatX to embrace the capital market anew.
From Buying a Chip Company to Competing for Custom Computing Power
If the $7 billion acquisition of MatX was a typical "vertical integration impulse," then after the negotiations failed, Anthropic clearly began to seriously consider another path: reshaping the computing power landscape through custom collaborations and limited self-development without swallowing an entire chip company. Reuters' report indicates that Anthropic is in talks with multiple chip startups, exploring self-developed or custom chip solutions to support the iterations of models like Claude. This means it is no longer trying to solve the problem with a single large acquisition but is breaking down "hardware capabilities" into multiple negotiable and combinable modules, distributing the technical and capital pressure among different partners. Some single-source information even claims that it is considering a multi-vendor, multi-chip strategy and investing several billion dollars in computing costs, but this claim remains to be verified, precisely reflecting its exploratory and complex approach to computing power planning.
The direct motivation driving this strategic assessment is not hard to understand. Currently, there is a high reliance on NVIDIA GPUs for large model training, and NVIDIA holds a significant dominant position in the related market; cost of computing power and supply constraints have long been recognized as pain points in the industry. As the Claude series rapidly iterates, Anthropic's training demands have expanded exponentially, and the high reliance on a single vendor has magnified risks in both cost and supply chain: prices have almost no room for negotiation, and queues and quotas during tight capacity can directly slow the pace of model releases. In this context, self-developed or custom chips are not just technical ideals but necessary "second choices" at the financial and operational levels, aiming to lower the unit cost of computing power for each training and inference while ensuring performance, allowing more room for product trial and error for Claude.
More critically, this crossroads choice is not an isolated event but is surrounded by a wave of self-development embodied by Google's TPUs and Amazon's Trainium. In recent years, cloud giants have transformed AI chips from "outsourced components" into "core capabilities written into the company fortress" through TPUs and Trainium, sending a message to the entire industry: in the era of large models, whoever controls the foundational computing power has the qualifications to plan the boundaries of products and ecosystems. The industry generally views chip capabilities as a part of the competitive advantage for large model companies, forcing players, including Anthropic, to consider whether and how to extend deeper into hardware. Rather than a high-risk large acquisition, Anthropic is currently choosing to engage in custom collaborations among multiple startups, exploring a "lightweight hardware path" that aligns more closely with its capital scale and technical rhythm amid the computing power torrent dominated by NVIDIA and the already established models of TPUs and Trainium.
MatX Without an Acquirer: Turning to Bet on $4 Billion Financing
For MatX, the moment Anthropic closed the door on the acquisition, the pressure of cash and time was thoroughly exposed. Subsequent reports from Reuters show that MatX quickly turned around and began seeking approximately $4 billion in new financing in the primary market, though it has not been clearly defined whether this figure represents fundraising goals or overall valuation. This deliberate ambiguity itself is like a test sheet gauging investors' appetites: the number is big enough but does not rush to explain whether it is "how much is needed" or "how much it is worth," first seeing if the market is willing to treat it as the next serious asset gamble.
From an industry inertia perspective, this level is not surprising. The development of high-performance AI chips and the validation of tape-out to build capacity each step burns tens of billions of dollars. For MatX, still in the critical stages of product development and market exploration, attempting to move forward independently without the safety net of acquisition can only view the capital it raises as its sole cushion. A $4 billion financing scale, if it is a fundraising goal, signifies that it aims to directly match the funding density of leading chip players; if it is a valuation coordinate, it announces to the market that it hopes to be categorized among "potential main suppliers" rather than "acquisition targets." Regardless of which interpretation, they point to the same reality: in the current scenario lacking clear major clients, product timelines, and public endorsements of technical parameters, the potential for MatX's technological iteration, capacity planning, and even entry into any mainstream computing power ecosystem will highly depend on whether this funding materializes and if capital is willing to treat it as the one that can truly deliver out of a batch of unproven AI chip stories.
The Subtle Trend of De-NVIDIA-ization: Multiple Startups Enter the Game
At the same time MatX attempts to use financing to reposition itself from "acquisition target" to "potential main supplier," another crucial clue disclosed by Reuters comes into play: Anthropic is not only focused on MatX but is also in talks with multiple AI chip startups, the identities of which have not been disclosed. This "one-to-many" approach itself is a de-NVIDIA-ization action—given NVIDIA's long-standing dominance in the AI training GPU market, which has nearly defined the rhythm of the computing power arms race, developers of large models like Claude are starting to proactively prepare a broader supply chain network for themselves. They are attempting to break free of the price, capacity, and technological route dependencies on a single supplier through custom or self-developed chips. Some sources even claim that Anthropic might combine multi-chip strategies among various suppliers, including NVIDIA and Google, and invest several billion dollars in computing costs for this purpose, but these claims have yet to receive broader verification, and their truthfulness remains to be validated further.
Once multi-supplier and custom chip strategies become mainstream choices in the industry, the layout at the bargaining table will be rewritten: with Google using TPUs and Amazon using Trainium internalizing critical hardware, the façade initially reserved for just cloud giants is now being explored by large model companies that attempt to pull some computing power logic back into their joint design with startups. For chip startups, being "invited to the table" by model players like Anthropic means they are no longer just waiting for orders to drop from cloud vendors but have opportunities to participate in the front lines of a new round of hardware-model collaborative evolution; however, the development and construction of high-performance AI chips still require tens of billions of dollars, and this opportunity also comes with intense capital constraints. For cloud vendors, self-developed chips and external startups are increasingly tied together by the same variable: whoever can find a new balance between cost control and technical discourse power is more likely to hold the initiative during the next reshuffling of the computing power landscape.
The Next Scene in the Computing Power Arms Race: Collaborative Game and Risk Reassessment
From once considering a $7 billion acquisition of MatX to publicly reported abandonment of the deal on August 28, 2026, while turning to negotiate self-developed or custom solutions with multiple chip startups, this step is not merely a tactical adjustment of a single company but a symbolic turning point in the computing power arms race: from "buying an entire factory" to "betting on multiple parallel options," control of computing power is being split into a more complex network of collaboration. As demand rises for large models like Claude, Anthropic and OpenAI are seen as core drivers of the NVIDIA GPU market, and now they are beginning to actively mitigate their reliance on a single supplier by placing self-developed chips, custom collaborations, and multi-supplier strategies on the same negotiating table. Correspondingly, MatX is seeking about $4 billion in financing after the acquisition falls through. The reality is that high-performance AI chips require tens of billions of dollars in funding, making capital in this arena role increasingly akin to a "second decision-making layer." In the future, it is more likely that there won't be blanket full acquisitions or complete outsourcing, but rather customized collaborations around specific models and scenarios, selective mergers of key teams and technologies, and a flood of capital entering chip startups, though this path remains fraught with informational blanks: the real considerations behind the acquisition termination have not been disclosed, the specific cooperation structures with other startups are unknown, and the results of MatX's financing have yet to see public conclusions. The next scene in the computing power arms race will unfold amidst the intertwining of collaborative games and risk reassessments.
Join our community, let's discuss together and grow stronger together!
AiCoin exclusive Hyperliquid benefits: https://app.hyperliquid.xyz/join/AICOIN88
AiCoin exclusive Aster benefits: https://www.asterdex.com/zh-CN/referral/9C50e2
On-chain Telegram community: https://t.me/AiCoinWhaleData
On-chain community: https://www.aicoin.com/link/chat?cid=N6OVMor5g
AiCoin on-chain Twitter: https://x.com/aicoinwhaledata
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
