Who’s buying AI chip startups?

Who’s actually buying AI chip startups right now?

The active buyers are mostly established semiconductor companies and strategic technology owners. AMD, Marvell, Qualcomm, Nvidia and SoftBank account for most of the visible dealmaking, while the largest cloud companies usually build chips internally or work with partners.

Across the recent transactions that most directly involve AI accelerators, specialized inference silicon, custom compute and the interconnect technology around those chips, the buyer list is surprisingly short. AMD recently agreed to buy specialized inference-chip startup Taalas. Marvell bought Celestial AI and XConn. Qualcomm completed its acquisition of Alphawave Semi. SoftBank took over Graphcore. Nvidia went much bigger with Groq, although that transaction used an unusual licensing-and-talent structure rather than a conventional takeover.

The money has moved well beyond acqui-hire territory. Four recent full acquisitions with usable public or reported values — Celestial AI, XConn, Alphawave and Graphcore — represent roughly $6.7 billion combined. Nvidia’s separate Groq arrangement was reported at about $20 billion on its own. Taalas adds another transaction whose price has not been disclosed.

Each buyer is plugging a different hole in its AI infrastructure. AMD wants more ways to run inference. Marvell wants the links connecting huge numbers of accelerators. Qualcomm wants data-center and edge-computing technology. Nvidia wants architectures that can strengthen its already dominant platform. SoftBank wants strategic ownership of next-generation compute through a portfolio that already includes Arm.

Buyer	Target	What the buyer wanted	Public or reported value	Structure
Nvidia	Groq	Specialized AI inference architecture and engineers	About $20B reported	IP licensing + key hires
AMD	Taalas	Highly specialized inference silicon	Undisclosed	Full acquisition announced
Marvell	Celestial AI	Optical scale-up interconnect	$3.25B upfront, potentially up to $5.5B	Full acquisition
Marvell	XConn	PCIe and CXL switching silicon	$469M purchase consideration	Full acquisition
Qualcomm	Alphawave Semi	High-speed connectivity and custom silicon	About $2.4B enterprise value	Full acquisition
SoftBank	Graphcore	AI accelerator technology and engineering team	About $600M reported	Full acquisition

If you want more recent data on this point, please see our latest AI chip market report.

What does “buying an AI chip startup” actually mean now?

Buying an AI chip startup today can mean taking the whole company, hiring the core engineering team, or paying for the technology while leaving the original company alive.

The distinction matters because some of the largest transactions would disappear from a traditional M&A database. Nvidia and Groq explicitly described their agreement as a non-exclusive technology license. Groq continued operating independently, while founder Jonathan Ross, president Sunny Madra and other employees moved to Nvidia. The reported price was around $20 billion, making the economic significance enormous even though Groq itself survived.

AMD used another version of this model with Untether AI in 2025. Untether had raised more than $150 million to build energy-efficient inference chips. AMD acquired its team through a strategic agreement and said those engineers would work on compiler development, kernels, digital and SoC design, verification and product integration.

SoftBank’s Graphcore purchase shows the cleaner version. Graphcore became a wholly owned SoftBank subsidiary, retained its name and continued developing AI computing technology.

A legal M&A screen therefore misses part of the market. In AI chips, a license plus a team transfer can move most of the strategic value while the company name stays alive.

Is Nvidia actually buying AI chip competitors?

Nvidia is willing to absorb strategically useful AI-chip technology, but the Groq transaction shows that it can get what it wants without purchasing every part of a rival company.

Groq had spent years developing its Language Processing Unit, or LPU, around a different approach to AI inference. Its pitch was extremely fast, predictable inference without relying on the same architecture as Nvidia GPUs.

Groq had also built meaningful developer adoption. Shortly before the Nvidia agreement, the company said its platform was being used by more than two million developers, up sharply from the previous year.

Nvidia took the technology and some of the people who understood it best. It did not sit on the asset for long: Nvidia subsequently introduced its Groq 3 LPX inference system as part of its broader AI infrastructure roadmap. Meanwhile, Groq raised new capital and pivoted toward operating AI infrastructure and neocloud capacity, including Nvidia systems. A strange deal, yes, but an economically huge one.

The price shows how seriously Nvidia valued the asset. A reported $20 billion is several times Groq’s previous $6.9 billion private valuation. The message is pretty clear: Nvidia will pay heavily when a different architecture can strengthen an important part of its AI-compute platform.

Is AMD the most aggressive buyer in AI chips right now?

AMD currently looks like the most systematic acquirer among Nvidia’s direct semiconductor rivals because it has been buying across AI silicon, software, engineering and complete systems instead of relying on one big transaction.

We counted seven important AI-related acquisition or team deals since 2023: Mipsology, Nod.ai, Silo AI, ZT Systems, Brium, the Untether AI team and Taalas. More revealing than the count is the spread across the stack.

AMD needed better software around its accelerators, so Nod.ai and Brium strengthened compilation and model optimization. Silo AI added hundreds of AI specialists and experience deploying models on AMD hardware. ZT Systems brought rack-level design and customer-enablement expertise. AMD then sold the manufacturing operation to Sanmina while keeping the design capabilities it actually wanted. The final purchase consideration reported for ZT Systems was about $4.4 billion. Silo AI had already cost another $665 million.

Hardware is now moving higher on the list. The Untether team brought inference-chip experience, and AMD recently agreed to acquire Taalas, a young company building extremely specialized silicon around individual AI models. Taalas had disclosed about $219 million of funding before the acquisition announcement.

Taken together, the sequence looks deliberate: software, model deployment, systems, then specialized inference hardware. AMD is using acquisitions to close several of the gaps separating a good accelerator from a complete AI platform.

AMD deal	Capability added	What it helps AMD do
Mipsology	AI inference software	Run models efficiently on AMD hardware
Nod.ai	Compilers and open-source AI software	Make AMD accelerators easier to program
Silo AI	AI models and deployment expertise	Help customers actually use AMD compute
ZT Systems	Rack and cluster design	Deliver complete AI systems faster
Brium	Compiler and inference optimization	Improve performance across the software stack
Untether AI team	AI-chip engineering	Strengthen inference, SoC and low-level software expertise
Taalas	Specialized inference silicon	Add a very different architecture for high-volume inference

If you want more recent data on this point, please see our latest AI chip market report.

Why is Marvell buying so many AI interconnect startups?

Marvell is currently one of the clearest buyers of AI-chip infrastructure startups because the company is betting that moving data between accelerators will become almost as valuable as the accelerators themselves.

Look at what Marvell bought during one short stretch. Celestial AI brought optical scale-up interconnect technology. XConn brought PCIe and CXL switching silicon. Polariton Technologies added high-speed, low-power plasmonics technology for future optical links. Those three deals attack different pieces of the same physical problem.

AI clusters keep getting larger. Once hundreds or thousands of accelerators need to work together, performance depends heavily on how quickly processors can exchange data and access memory. Adding faster chips becomes less useful when those chips spend too much time waiting for information to arrive. Electrical connections also become harder to scale as bandwidth, distance and power requirements rise.

That explains why Marvell paid $3.25 billion upfront for Celestial AI, with another $2.25 billion potentially available if the business reaches ambitious revenue milestones. It paid $469 million of final purchase consideration for XConn. Polariton’s price was undisclosed.

Instead of spending billions on another general-purpose accelerator, Marvell is buying the infrastructure needed to connect accelerators at scale, regardless of which processor architecture ultimately wins.

Why is Qualcomm suddenly buying so much AI-chip technology?

Qualcomm is buying AI-chip technology because the company wants a much bigger business in data centers and edge AI than smartphones alone can give it.

The $2.4 billion Alphawave Semi acquisition is the clearest data-center move. Alphawave develops high-speed wired connectivity, chiplets and custom silicon. Qualcomm explicitly tied the transaction to its expansion into data centers, where its Oryon CPUs and Hexagon NPUs need fast connectivity if the company wants to compete for serious AI workloads.

Qualcomm has followed this playbook before. It bought Nuvia for about $1.4 billion in 2021, gaining a team of former Apple chip engineers and the CPU technology that eventually became central to Qualcomm’s Oryon roadmap. Nuvia was only around two years old when Qualcomm agreed to buy it.

The other half of Qualcomm’s strategy sits at the edge. Qualcomm said earlier this year that five acquisitions completed over an 18-month period — Augentix, Arduino, Edge Impulse, FocusAI and Foundries.io — had helped reshape its industrial and embedded IoT business. Edge Impulse alone brought a platform used by more than 170,000 developers to build and deploy AI models on embedded hardware.

Qualcomm’s logic is straightforward. It buys heavy silicon and IP where it wants to enter large compute markets, while software and developer acquisitions make its existing chips easier to use in cameras, robots, industrial equipment and other edge devices.

That makes Qualcomm a more credible future buyer than its old reputation as a smartphone-chip company suggests. A startup with efficient AI compute, high-speed connectivity or technology that makes on-device models easier to deploy now fits directly into several Qualcomm businesses.

Why aren’t Google, Amazon, Microsoft and Meta buying more AI chip startups?

Google, Amazon, Microsoft and Meta are huge users of AI chips, but today these companies behave more like chip builders and strategic partners than serial buyers of independent AI-chip startups.

Amazon is the clearest example of how one acquisition can remove the need for repeated deals. AWS bought Annapurna Labs in 2015 and turned that team into the foundation for Nitro, Graviton, Inferentia and Trainium. According to Amazon’s recent update, its custom-chip business has now exceeded a $25 billion annual revenue run rate and is growing at triple-digit percentages year over year.

Google took an even more organic path. Its TPU program grew from an internal custom-silicon effort that began more than a decade ago. Google says more than 100,000 first-generation TPUs were eventually deployed after engineers initially expected fewer than 10,000. These days Google combines that internal design expertise with external partners. Its latest expanded Marvell relationship covers custom accelerators, memory controllers and other silicon, with warrants allowing Google to buy up to roughly $12.2 billion of Marvell shares if milestones are met.

Microsoft has used selective M&A. It acquired Fungible in 2023 to bring its data-processing-unit technology and engineering team into Azure’s data-center infrastructure organization. Meta has leaned much harder on internal MTIA chips and co-development. Meta currently plans four new MTIA generations over two years and has expanded its Broadcom relationship into a multi-generation custom-silicon program beginning with more than one gigawatt of deployment.

For AI-chip founders, hyperscalers are less obvious exit buyers than their enormous chip budgets suggest. These companies can buy a startup when it saves years of work, but most of the time they have enough scale to hire, build or co-develop instead. Their spending power is real; their need to acquire is much smaller.

Company	Current custom-chip approach	Important acquisition history	What it suggests for startups
Amazon	Large internal Annapurna organization	Annapurna Labs	One successful acquisition became a long-term in-house chip engine
Google	Internal TPU design + external co-development	No comparable recent accelerator takeover	Partnership can be more attractive than buying the supplier
Microsoft	Internal silicon + selective acquisitions	Fungible	Buys when a specific infrastructure capability can accelerate Azure
Meta	Internal MTIA + Broadcom co-development	No major recent AI-accelerator takeover	Prefers internal design and strategic suppliers

If you want more recent data on this point, please see our latest AI chip market report.

Are buyers paying for AI chip revenue, or buying the technology years early?

AI-chip buyers are currently willing to pay for architecture years before the acquired business reaches meaningful revenue, especially when that technology could remove a future infrastructure bottleneck.

Celestial AI makes the case unusually clear. Marvell agreed to $3.25 billion of upfront consideration even though Marvell expects Celestial’s initial revenue contribution only later in its planning horizon. Marvell projects a $500 million annualized revenue run rate by the fourth quarter of its fiscal 2028 and $1 billion a year later. The earnout itself is tied to Celestial eventually producing substantial revenue.

Nuvia followed a similar logic years earlier. Qualcomm’s SEC filings described the company as having in-process technologies and said Nuvia’s operating results were not material at the time. Qualcomm still paid roughly $1.4 billion because it wanted the CPU design team and roadmap.

AMD’s Taalas deal follows the same pattern. Taalas had built a highly specialized architecture and raised around $219 million, but AMD did not wait for it to become a large independent semiconductor supplier before moving.

Revenue can come later. What buyers are often paying for is engineering time: patents, compiler expertise, tape-outs and a team capable of producing the next chip can justify prices that current sales alone would never explain.

Are AI chip acquisitions creating great exits or rescue exits?

AI-chip acquisitions are producing both spectacular outcomes and painful resets, with a startup’s bargaining power changing dramatically once it either proves a unique architecture or starts running out of capital.

Graphcore is the cautionary case. The British AI-chip company reached a private valuation of about $2.8 billion and raised roughly $700 million, yet struggled to turn strong technology into enough commercial adoption. Financial reporting later showed only $2.7 million of sales against a $205 million pre-tax loss in 2022. SoftBank eventually bought the company for an undisclosed amount that the Financial Times reported at slightly above $600 million.

Using that reported figure, the exit came almost 80% below Graphcore’s old peak valuation and below the total venture capital invested. The technology survived and Graphcore gained a deep-pocketed owner, but earlier investors were looking at a very different outcome from the one implied by the company’s peak valuation.

Nuvia represents the opposite kind of transaction. The company announced a $53 million Series A and a $240 million Series B before Qualcomm agreed to acquire it for about $1.4 billion. The purchase price was roughly 4.8 times those two announced funding rounds combined. That is not an investor-return calculation, since ownership percentages and other financing details matter, but it shows how differently buyers can price scarce chip talent.

Semiconductor economics create that spread. A startup can burn hundreds of millions before proving demand. If customers arrive while there is still plenty of cash in the bank, the founders negotiate from strength. If another funding round is needed just to keep the roadmap alive, the buyer may control the negotiation.

If you want more recent data on this point, please see our latest AI chip market report.

Is China buying AI chip startups the same way as the US?

China’s AI-chip market currently has plenty of semiconductor consolidation, but whole-company accelerator acquisitions play a smaller role than domestic investment, procurement and broader industrial consolidation.

One recent transaction shows what Chinese strategic M&A can look like. PATEO agreed to buy roughly 70% of Chengdu Mingyi for up to RMB1.4 billion. Chengdu Mingyi develops high-speed optoelectronic and analog chips, including technology used in 400G and 800G optical modules for AI data centers. PATEO sees that technology as part of a broader combination of software, hardware, chips and cloud infrastructure.

Much larger consolidation is happening around manufacturing. SMIC has been moving to take full control of a major foundry subsidiary in a transaction worth several billion dollars, while Hua Hong has pursued its own billion-dollar foundry consolidation. Those deals strengthen the semiconductor supply chain rather than absorbing independent AI-accelerator startups.

Chinese internet companies are also behaving differently from US semiconductor buyers. ByteDance has recently been discussing purchases of inference chips from Iluvatar CoreX and Baidu’s Kunlunxin while already using Huawei and Cambricon as domestic suppliers. Alibaba has been putting more capital into its own T-Head chip operation. Procurement and internal development can achieve the strategic goal without buying the supplier.

Geopolitics reinforces that behavior. Chinese companies need domestic alternatives to restricted Nvidia products, and Beijing wants multiple local semiconductor capabilities to grow. Buying every promising startup into one incumbent would work against some of that ecosystem-building goal.

The result is a chip M&A market with fewer headline acquisitions resembling AMD buying a Western inference startup. The money is spread across factories, strategic stakes, internal chip units, suppliers and selective takeovers. That is a very different route from the US playbook.

What kind of AI chip startup is easiest to sell today, and which ones have become too expensive?

The easiest AI-chip startups to sell today are specialists that remove a painful bottleneck at a price a strategic buyer can still digest; the biggest independent accelerator companies are increasingly too expensive for routine acquisitions.

Recent connectivity deals make the first part clear. Marvell bought Polariton for high-speed optical modulation technology. Credo acquired DustPhotonics to add silicon-photonics integrated circuits. Molex acquired Teramount for fiber-to-chip connectivity aimed at co-packaged optics. XConn fits the same pattern on the electrical side, with PCIe and CXL switching technology that Marvell could plug directly into its networking roadmap.

These companies are easier to integrate than a general-purpose accelerator challenger. A buyer knows where the technology fits, which products it improves and which customers can buy it. Acquiring a full accelerator startup often means inheriting the architecture, compiler, software libraries, manufacturing roadmap, systems engineering and developer ecosystem together.

At the same time, the strongest accelerator companies are becoming extremely expensive. Etched has recently been valued at about $21 billion after raising nearly $2 billion in total capital and reportedly securing more than $1 billion of orders. SambaNova’s latest $1 billion financing valued it at $11 billion. d-Matrix reached a $2 billion valuation after raising $450 million. Cerebras has gone public and its recent market capitalization has been around $46 billion to $50 billion.

Compare those figures with Qualcomm’s roughly $2.4 billion Alphawave deal. Etched’s latest private valuation is almost nine times larger, SambaNova’s is more than four times larger and Cerebras trades at close to twenty times it.

That price gap explains much of the current acquisition pattern. Buyers can still spend hundreds of millions or a few billion dollars on inference technology, optics, interconnect, memory movement and software. Buying one of the leading independent accelerator companies now requires a board-level bet measured in tens of billions.

If you want more recent data on this point, please see our latest AI chip market report.

So, who’s buying AI chip startups?

AI chip startups are being bought mainly by semiconductor incumbents and strategic infrastructure owners. AMD is the broadest buyer, Marvell is the most concentrated buyer of AI connectivity technology, Qualcomm is a serious data-center and edge buyer, Nvidia is a highly selective IP-and-talent buyer, and SoftBank remains an opportunistic strategic owner.

The surprising part is who does not dominate the list. Google, Amazon, Microsoft and Meta spend staggering amounts on AI hardware, yet all four have strong reasons to build chips internally or co-develop them with established semiconductor suppliers. A startup should not assume that the company buying millions of accelerators will also want to buy the accelerator company.

The target mix has also changed. Buyers are increasingly interested in inference architectures, chip-to-chip communication, PCIe and CXL switching, silicon photonics, compilers and complete systems. As AI infrastructure grows from individual GPUs into multi-rack machines consuming enormous amounts of power, value is spreading into every technology that keeps those machines fed with data.

AMD and Marvell are especially important to watch. AMD has been assembling missing pieces across almost the whole AI stack, while Marvell has concentrated several acquisitions around the connectivity bottleneck. Qualcomm’s expansion gives another group of founders a realistic strategic buyer across both data centers and edge AI.

Nvidia deserves its own category. The company has enough market power to structure a deal around the exact technology and people it wants, even when that means leaving the original company operating independently.

The market is real and increasingly valuable, but much of the action has shifted one layer away from the obvious Nvidia challengers. Specialized inference, interconnect, optics, memory movement, systems and software are where strategic buyers can still find technology that matters at prices they can realistically pay.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.