Nvidia's $40 Billion Arm Deal and How Regulators Killed It

On September 13, 2020, Nvidia agreed to buy Arm Limited from SoftBank Group for $40 billion in cash and stock, the largest acquisition ever attempted in the semiconductor industry. Seventeen months later, on February 8, 2022, the two companies walked away. No court had ordered them to stop. No regulator had issued a final prohibition. The deal died because four competition authorities on three continents had each opened an investigation, none of them was close to finishing, and the buyer concluded that the calendar had beaten it.

That is the first thing that makes this case unusual. The second is what happened next. SoftBank kept $1.25 billion of a non-refundable prepayment, floated Arm on Nasdaq in September 2023 at roughly $54.5 billion, and watched the public market eventually value the company at several times what Nvidia had agreed to pay. Nvidia, meanwhile, built the Arm-based data center processor it wanted under a license it already held, and went on to become the most valuable company in the world.

The question worth working through is therefore not simply why the deal failed. It is whether the failure destroyed value or merely reallocated it, and whether the neutrality argument the regulators accepted was a genuine competition problem or a successful lobbying campaign by Arm's largest customers. This study reconstructs the transaction from the parties' own filings and releases, the regulators' published decisions, and the analysts who argued with each of them at the time.

Every failed deal has a seller who was willing, and SoftBank's willingness is the part of this story most often skipped. Masayoshi Son had owned Arm for barely four years when he agreed to sell it, at a price that valued the asset only modestly above what he had paid.

The Twenty-Four Billion Pound Bet That Did Not Compound

SoftBank acquired Arm Holdings plc in 2016 for £24.3 billion, roughly $32 billion at the exchange rates of the day, then the largest foreign takeover ever completed by a Japanese company. The timing was opportunistic: sterling had collapsed after the Brexit referendum under four weeks earlier, and Son moved on a UK asset that had suddenly repriced in yen terms. His stated thesis was the Internet of Things, the idea that Arm's low-power designs would end up inside every connected appliance, sensor and vehicle.

The thesis was directionally right and commercially slow. SoftBank spent heavily on Arm's headcount and research programs, deliberately trading near-term profit for a larger long-run footprint, and Arm's revenue in the fiscal year ended March 31, 2021 was $2.03 billion, per the company's later F-1 registration statement. Four years of private ownership had produced a bigger engineering organization and a broader product roadmap, but not the compounding financial performance a $32 billion purchase price implied.

Arm's licensing model

Arm does not manufacture chips or sell them. It licenses processor designs and the underlying instruction set to semiconductor companies, which pay an upfront license fee and then a per-unit royalty on every chip shipped. Because Arm's customers compete with each other, the model only works if Arm is indifferent between them. That structural neutrality, not any single design, is the asset that the entire dispute over this deal was about.

The Vision Fund Hole and the Asset Sale Program

The immediate pressure came from somewhere else entirely. In March 2020, with the Vision Fund carrying heavy writedowns on WeWork and Uber and SoftBank's own shares under stress, the group announced a program to sell up to $41 billion of assets, directing $18 billion to buybacks and the remaining $23 billion to debt reduction and cash reserves. Most of that was expected to come from SoftBank's Alibaba position.

Arm was not originally on the list. It became a candidate because it was the one wholly controlled operating asset with a plausible strategic buyer at a price that would move the needle. The $2 billion Nvidia paid at signing was itself immediate liquidity, arriving in a fiscal year when SoftBank needed it.

The structure of the ownership complicated matters. SoftBank had agreed in 2017 to sell a 25% stake in Arm to the first Vision Fund for $8.2 billion, transferred in a March 2018 reorganization, which implied a valuation just under $33 billion and meant that any sale had to clear a price at which the fund's limited partners were not visibly impaired. A $40 billion headline cleared that bar, but not by much.

Why a Trade Sale Beat an IPO in 2020

An initial public offering was always the alternative exit, and SoftBank had considered it before. In 2020 it looked poor. Arm's revenue had been broadly flat across the SoftBank years, its Chinese joint venture was in open governance revolt, and the equity story rested on data center and automotive royalties that had not yet shown up in the numbers. A listing would have been priced off trailing financials that told an unflattering story.

A strategic sale solved that problem by selling the future rather than the past. Nvidia was buying an ecosystem thesis, not a $2 billion revenue line, and it was willing to pay in a currency (its own shares) that it believed was undervalued. The counterfactual matters here: had SoftBank listed Arm in 2020 on the same multiple the market eventually applied in 2023, it would have raised a fraction of what the Nvidia agreement promised. The decision to sell was rational on the information available, and it looks wrong only because of what happened to semiconductor valuations afterward.

Nvidia's side of the logic was industrial rather than financial. The company had spent a decade turning the graphics processor into the default engine of machine learning, and by 2020 it had a gap in its product line that its competitors did not.

What Nvidia Owned, and the One Thing It Did Not

Nvidia sold accelerators. Every accelerator sat next to a CPU it did not make, usually an x86 part from Intel or AMD, and increasingly the CPU rather than the accelerator set the ceiling on system performance. Nvidia had already bought Mellanox to own the networking layer between servers. The processor was the remaining piece of a full data center platform.

Arm was the only realistic way to acquire a general-purpose CPU architecture. x86 was licensed to a closed set of incumbents; RISC-V was open but immature; designing a new instruction set from nothing would have meant building a software ecosystem for a decade. Arm came with an installed base that Nvidia's own announcement put at 180 billion chips shipped and more than 1,000 technology partners.

Instruction set architecture

The instruction set architecture, or ISA, is the contract between software and silicon: the vocabulary of operations a processor understands, and therefore the thing every compiler, operating system and application is written against. Owning a widely adopted ISA is closer to owning a language than to owning a product, because the switching cost for customers is the cost of recompiling and revalidating everything they have ever shipped. Arm's ISA and x86 are the two that matter at scale, with RISC-V the open-source challenger.

The License That Made the Purchase Optional

The most awkward fact in Nvidia's case for the deal is that it did not need to own Arm to design Arm chips. Nvidia held an architecture license, the highest tier of Arm agreement, which permits a licensee to design its own core implementing Arm's instruction set. On April 12, 2021, seven months into the regulatory review and with the acquisition still pending, Nvidia announced Grace, its first Arm-based data center CPU, built on licensed Arm IP under that existing license.

Grace demonstrated the point that regulators and rivals would both seize on. If the license was sufficient to build the product, the acquisition was buying something else: control of the roadmap, of pricing, and of which competitors got early access to which technology. Nvidia's answer was that owning Arm would let it push accelerated computing into the ecosystem far faster than a licensee ever could, and that the developer reach was the prize. Jensen Huang put the arithmetic in his letter to employees, saying the deal would expand Nvidia's developer reach from 2 million to more than 15 million, and promised in the same letter that Nvidia would "maintain its open-licensing model and customer neutrality."

The Promises Attached to the Cambridge Address

Nvidia understood from the first day that this was a political transaction as much as a commercial one, and the announcement was loaded with commitments designed to pre-empt a UK backlash. Arm would keep its name and its Cambridge headquarters, its intellectual property would stay registered in the United Kingdom, and Nvidia would build a new AI research center on the Cambridge campus.

Those commitments addressed the wrong objection. The UK's eventual concern was not where Arm's staff sat but whether an Arm owned by a chip designer would keep treating that designer's competitors as equals. Promises about buildings and headcount are easy to verify and easy to honor; promises about the future direction of a research roadmap are neither. That asymmetry is what the Competition and Markets Authority would later build its rejection of Nvidia's remedies on.

The consideration structure repays close reading, because almost every feature of it was doing work: managing Nvidia's cash outlay, giving SoftBank exposure to Nvidia's upside, and pricing the risk that regulators would kill the deal.

The Consideration Stack, Piece by Piece

Nvidia agreed to pay SoftBank $12 billion in cash, of which $2 billion was payable at signing, and to issue 44.3 million Nvidia shares with an aggregate value of $21.5 billion measured against a 30-day average closing price. SoftBank could earn up to a further $5 billion, in cash or in as many as 10.3 million additional Nvidia shares, if Arm hit specified financial targets in the fiscal year ending March 31, 2022. A separate $1.5 billion of Nvidia equity was earmarked for Arm employees, and Arm's IoT Services Group was carved out and retained by SoftBank.

The table reconciles to the headline figure, but the headline figure overstates what SoftBank was contracted to receive. Only $33.5 billion was firm consideration to the seller; $5 billion was contingent and $1.5 billion never belonged to SoftBank at all. Deal headlines routinely bundle retention equity and earnouts into the announced number, and a candidate who can strip a press release back to the firm cash-and-stock component is doing something most commentary does not.

Earnout

An earnout makes part of the purchase price contingent on the target hitting agreed post-signing performance thresholds. It bridges a valuation gap: the buyer refuses to pay today for growth it cannot verify, and the seller refuses to hand over the upside for free. The mechanics, negotiation traps and accounting are covered in our explainer on earnout structures in M&A. Here the trigger was Arm's own financial performance in the year to March 2022, a period that expired while the deal was still stuck in review.

The Two Billion Dollars That Started a Clock

The signing payment is the structural feature that separates this deal from a conventional merger agreement. Nvidia wired $2 billion on day one, of which $1.25 billion was non-refundable if the transaction failed for the reasons it eventually failed for. In exchange, Nvidia locked up an asset with an expected close roughly 18 months away and secured a long-dated Arm license alongside it.

Read commercially, the prepayment was a reverse break fee paid up front rather than at the end. It gave SoftBank certainty of receiving something regardless of outcome, which mattered enormously to a seller with an immediate liquidity need, and it gave Nvidia a signed deal without a competitive process. Read as risk pricing, $1.25 billion against $40 billion is roughly 3%, in the normal range for an antitrust reverse fee on a deal with obvious regulatory exposure, and arguably light given that four jurisdictions were involved from the outset.

Break fee

A break fee, or termination fee, is a contractually fixed payment owed when a signed deal does not close. A reverse break fee runs the other way, from buyer to seller, and typically compensates the target for the time its business spent frozen inside a failed process. The full mechanics are set out in our guide to break-up and termination fees. What is unusual here is the timing: the money moved at signing, not at termination, so SoftBank never had to sue anyone to collect it.

Why a Fixed Share Count Made the Price Move

Nvidia agreed to deliver a fixed number of shares, 44.3 million, not a fixed dollar value. That is a fundamentally different bargain from the all-cash structures common in large strategic acquisitions, and it handed SoftBank the whole of Nvidia's equity performance between signing and closing.

Nvidia's shares then did what they did. The company finished calendar 2020 with a market capitalization of roughly $323 billion and calendar 2021 at roughly $736 billion, more than doubling while the deal sat in review. By the time the transaction was abandoned, the consideration was widely reported as worth more than $60 billion rather than the $40 billion announced. Nothing had been renegotiated; the currency had simply appreciated.

Opposition to this deal did not come primarily from consumer groups or from antitrust economists. It came from Arm's own customers, and it arrived within days of the announcement.

Hermann Hauser and the Sovereignty Argument

Hermann Hauser, who co-founded Arm in 1990, launched a public campaign against the sale the week it was announced, publishing an open letter to Prime Minister Boris Johnson and a petition site under the name Save Arm. His case had two strands. The commercial strand was that Nvidia, as a chip designer competing with Arm's licensees, would destroy the neutral licensing model on which several hundred customers depended.

The second strand was national. Hauser argued that placing Arm under US ownership would subject its export decisions to Washington, at a moment when US restrictions on Chinese technology companies were tightening rapidly.

That framing proved durable because it gave the UK government a hook that competition law alone did not provide. The Secretary of State would eventually intervene on national security grounds, not only on competition grounds, and the two reviews were then run together.

Why Qualcomm, Microsoft and Google Complained in Four Places

By February 2021, Qualcomm, Microsoft and Google had each taken their concerns to antitrust authorities. Qualcomm's objection was the most explicit and the most consequential: it told the FTC, the European Commission, the CMA and China's State Administration for Market Regulation that Nvidia would become a gatekeeper of Arm's technology able to restrict rivals' access to the designs their entire product lines were built on.

The choice to complain in four jurisdictions simultaneously was a strategic one, and it is the single most important tactical fact in the case. A merger of this shape needs every clearance; it needs only one refusal to die. Complainants therefore do not pick the regulator most likely to agree with them, they brief all of them, because the probability that at least one authority opens an in-depth review rises sharply with the number of doors knocked on. The same asymmetry that shaped Microsoft's fight over Activision Blizzard, where a single UK decision nearly killed a deal cleared elsewhere, applied here with four regulators instead of three.

The customers' underlying commercial interest was straightforward and was never really disguised. Arm licensees pay royalties to a company that has no reason to prefer any one of them. An Arm owned by Nvidia would have had a reason. Whether that reason would ever have been acted on is the question the regulators had to answer, and neither the licensees nor Nvidia could prove their case in advance.

Six Undertakings and a Monitoring Trustee

Nvidia's remedy package, offered to the CMA during the first phase of the UK review, was a set of behavioral commitments: preserve Arm's open licensing program for five years, provide equal access to all licensees, grant early access to architectural licensees, deliver IP without interoperability restrictions, honor confidentiality obligations toward licensees' sensitive information, and submit to an independent monitoring trustee.

The CMA rejected the package on two grounds that are worth separating. The first was duration. The regulator found that a five-year commitment falls "manifestly short of the time period required to remedy the concerns identified", which it described as lasting in nature. The second was observability: even with a trustee in place, the CMA noted an inherent reliance on third parties to notice and report suspected breaches. Neither objection was about Nvidia's good faith. Both were about whether anyone could tell, from outside, if a research roadmap had been quietly tilted.

Thirteen months separated the CMA's first inquiry from the termination, and across four authorities no two of them were reviewing quite the same question. Understanding the sequence is the fastest way to see why the deal became undeliverable long before anyone formally prohibited it.

London: From Public Interest Notice to Phase 2

The CMA opened its file on January 6, 2021 and never closed it in Nvidia's favor. Its phase 1 work, delivered to the Secretary of State on July 20, 2021 and published in summary a month later, found a realistic prospect of a substantial lessening of competition across six markets grouping into four areas where Nvidia and Arm licensees met: data centers, the Internet of Things, automotive, and gaming. The mechanism was input foreclosure. Nvidia would control an input its competitors could not replace and would have both the ability and the incentive to degrade their access to it.

CMA chief executive Andrea Coscelli framed it plainly, saying the regulator was concerned that Nvidia controlling Arm could create real problems for Nvidia's rivals by limiting access to key technologies. The novelty was that this was a vertical case. Nvidia and Arm were not competitors in any conventional market-share sense, and for most of the previous four decades a vertical merger with these facts would have attracted remedies rather than prohibition.

The UK escalation on November 16, 2021 was formally a government decision rather than a regulator's. Digital Secretary Nadine Dorries referred the transaction to a full phase 2 investigation on both competition and national security grounds, noting that Arm has "a unique place in the global technology supply chain". The reference started a 24-week clock, extendable by eight weeks, that would have run into mid-2022 at the earliest.

Phase 2 reference

Under the UK merger regime, the CMA conducts a short phase 1 screen and refers a transaction to phase 2 when it finds a realistic prospect of a substantial lessening of competition. Phase 2 is a full inquiry before an independent panel with the power to prohibit the deal or impose structural remedies, and it typically runs 24 weeks. A reference is not a decision, but for a buyer working against a contractual outside date it is often decisive, because the time cost of losing is as damaging as the loss.

Brussels: The Information-Flow Theory

The European Commission opened its in-depth phase II investigation on October 27, 2021, with a provisional decision deadline of March 15, 2022. Its theory overlapped with the CMA's on foreclosure but added a distinct concern that has since become a standard feature of vertical technology cases.

Arm's licensees share confidential roadmap information with Arm as a matter of routine, because a chip design cannot be supported without knowing what the customer is building. If the licensor is owned by a competitor, that flow of information becomes hazardous, and the Commission's worry was that licensees would simply stop sharing. The harm in that scenario is not foreclosure by refusal; it is a slow degradation of the technical collaboration on which the whole ecosystem runs.

Brussels never reached a decision. The investigation was still running when the parties terminated, and the file was closed without a published finding on the merits. That absence matters when assessing the case: the EU's view of this transaction is unresolved, not adverse.

Washington: The FTC's Three Markets

The most concrete legal action came from the United States. On December 2, 2021 the Federal Trade Commission voted 4-0 to issue an administrative complaint seeking to block the deal, and set an administrative trial for August 9, 2022. The unanimity is notable: this was not a divided commission stretching doctrine, it was every sitting commissioner agreeing.

The complaint identified three worldwide markets in which Nvidia sold Arm-based products and could foreclose rivals who needed Arm technology: processors for high-level advanced driver assistance systems in passenger cars, DPU SmartNICs used to offload networking and security work in data center servers, and Arm-based CPUs for cloud computing providers. It added the information-access theory, alleging Nvidia would gain visibility into the competitively sensitive plans of licensees who were also its rivals.

Two procedural points shaped the outcome. The FTC chose its in-house administrative forum rather than seeking a federal court injunction, which meant a slower path but no early judicial test of the theory. And it had withdrawn its own Vertical Merger Guidelines in September 2021, signaling that the permissive framework under which deals of this shape had been cleared for years was no longer operative. For candidates working through how these reviews actually run, our primer on antitrust and regulatory approval in M&A covers the process mechanics.

Beijing: The Approval That Never Started

China was the jurisdiction nobody could solve. Nvidia filed the transaction with the State Administration for Market Regulation in 2021, but SAMR did not accept the formal notification until January 2022, three weeks before the parties walked away, so the substantive review clock had barely begun to run.

The political context made a Chinese clearance close to impossible. Arm-based designs sit at the center of Chinese semiconductor self-sufficiency plans, US export controls on Chinese technology firms were tightening through the entire review period, and approving the transfer of the underlying architecture to a US owner would have run directly against Beijing's strategic interest. Regulatory silence, in that setting, is a decision.

Underneath the politics sat a governance disaster that made the Chinese position even harder to negotiate. Arm China, the joint venture in which Arm had sold 51% to a consortium of local investors for $775 million, was effectively outside Arm's control. Its chief executive Allen Wu had been voted out by the board in June 2020 over undisclosed conflicts of interest, refused to leave, and retained practical control because he held the company seal that Chinese law treats as the binding signature of the entity. The dispute was not resolved until Allen Wu was finally displaced in the spring of 2022, months after the acquisition had already collapsed, and by then Arm had sold its entire Arm China equity stake to a SoftBank subsidiary, keeping an indirect interest of under 5% and no board rights. A buyer cannot give a regulator comfort about an asset the seller does not command.

The table makes the central point visible. Not one of these authorities had blocked the deal on February 8, 2022. Three had unresolved in-depth proceedings and one had not started, and the earliest realistic path to a full set of clearances ran well past the transaction's expected close. That is the specific way modern cross-border mega-deals die: not by prohibition, but by the accumulation of parallel timetables that no buyer can outlast. The same dynamic is examined more generally in our note on cross-border M&A considerations.

Master the mechanics behind a deal answer: practice 1,000+ technical questions on M&A, financing, and accretion/dilution, download our iOS app for the full toolkit.

The collapse itself was quick, and the sequence tells you who was managing the narrative.

The January Leak

On January 25, 2022, Bloomberg reported that Nvidia was quietly preparing to abandon the acquisition, having told partners it did not expect the transaction to close, and that SoftBank was moving ahead with preparations for an Arm IPO. Neither company confirmed it. Two weeks later the report was correct in every particular.

Leaks of this kind are rarely accidental. Signaling an imminent walk-away lets both sides manage their own shareholders before the announcement, and it lets the seller start marketing the alternative exit while the buyer absorbs the failure. By the time the formal statement came, the market had already repriced the outcome.

February 8, and What Each Side Kept

The joint announcement on February 8, 2022 attributed the termination to significant regulatory challenges preventing consummation of the transaction. SoftBank confirmed that it would retain the $1.25 billion non-refundable portion of the signing payment, recording it as profit in the quarter ended March 31, 2022. Nvidia retained the 20-year Arm architecture license that the same signing payment had also purchased.

Nvidia's accounting told a slightly different story from SoftBank's. The company recorded an acquisition termination charge of $1.35 billion in the first quarter of its fiscal 2023, writing off the prepayment provided at signing, an after-tax impact of $0.52 per diluted share in its reported results. The two figures measure different things. Nvidia had allocated $1.36 billion of the $2 billion to advance consideration for the acquisition and the remaining $0.64 billion to a 20-year prepayment of Arm licenses and royalties, so the write-off is of its own capitalized prepayment balance for the deal, not of the $1.25 billion deposit SoftBank booked as profit.

Rene Haas and the Immediate Pivot

On the same day the termination was announced, Arm replaced its chief executive. Simon Segars, who had been at the company for 30 years and had led it through the SoftBank ownership and the failed sale, stepped down. Rene Haas was appointed CEO with immediate effect, having joined Arm in 2013 after running Nvidia's own IP business.

The choice signaled the new destination. Haas had run Arm's IP Products Group, the commercial heart of the licensing business, and his mandate was to prepare a company that had never reported to public shareholders for a listing. SoftBank said it would begin preparations for a public offering within the fiscal year ending March 31, 2023, which set a timetable it very nearly met.

Nineteen months after the termination, the market delivered its own verdict on what Arm was worth, and it landed remarkably close to what Nvidia had agreed to pay.

Buying Back the Vision Fund's Quarter at Sixty-Four Billion

Before the listing SoftBank had to tidy the cap table. In August 2023 it acquired the Vision Fund's 25% stake in Arm at a valuation of roughly $64 billion, taking its ownership to 100% ahead of the offering. The stated rationale was to remove an overhang, since the fund had planned to sell down into the public market over time, whereas SoftBank intended to hold.

The transaction has an awkward feature that is worth naming. SoftBank Group was on both sides: it bought from a fund it manages, at a valuation about 17% above the price at which the same asset would be sold to public investors three weeks later. The Vision Fund's limited partners were the beneficiaries of the higher mark, and SoftBank Group's own balance sheet paid for it. Related-party pricing inside a group structure is legitimate and disclosed, but it is not a market price, and the IPO immediately showed the difference.

Pricing at Fifty-One Dollars, and the Book of Customers

Arm priced its offering on September 13, 2023 at $51.00 per American depositary share, selling 95.5 million ADSs for gross proceeds of approximately $4.87 billion and valuing the company at roughly $54.5 billion. Every share sold was secondary: the seller was a wholly owned SoftBank subsidiary, and Arm itself raised nothing. Barclays, Goldman Sachs, J.P. Morgan and Mizuho ran the books.

The cornerstone list was the most interesting document in the deal. Apple, Nvidia, Alphabet, AMD, Intel and Samsung, several of them among the customers who had objected to Nvidia owning Arm, took allocations in the offering. Shares opened at $56.10 and closed the first day at $63.59, up close to 25% on the issue price. Our walkthrough of the IPO process covers how cornerstone books and pricing decisions of this kind are constructed.

There is an elegant symmetry in that investor list. The same firms that had lobbied four regulators to prevent Arm from falling under a single competitor's control then paid to become minority shareholders in an Arm that answered to no competitor. Distributed customer ownership of a neutral supplier is a different thing from ownership by one of its rivals, and the cornerstone book was the clearest available evidence that the neutrality objection had been about structure rather than about Nvidia specifically.

What the Public Market Said About the Forty Billion Price

The clean way to test whether Nvidia overpaid is to compare implied multiples at each repricing of the same asset.

The first two rows are the finding. On a revenue multiple, the price Nvidia negotiated privately in 2020 and the price public investors set in 2023 are effectively identical, which is unusual and analytically useful: it suggests the $40 billion was neither a strategic premium nor a distressed discount, but a fair market-clearing valuation for Arm as it then stood. The revenue figures come from Arm's F-1 and its fiscal 2026 results, and the fiscal year in the first row ended six months after the deal was signed.

The third row is a different animal entirely. Arm reported record revenue of $4.92 billion for the year to March 31, 2026, up 23%, with royalties of $2.61 billion and licensing of $2.31 billion. The multiple expansion from roughly 20 times to roughly 52 times revenue is not an Arm story so much as an AI-infrastructure story, and it is the reason the failed deal looks in hindsight like the transaction of the decade for the seller.

Go deeper before an interview: our 160-page PDF covers deal frameworks, technicals, and behavioral prep end to end, access the IB Interview Guide and pair it with these case studies.

The counterfactual is unusually easy to observe here, because Nvidia went ahead and executed the underlying strategy without the acquisition.

Grace, and the License That Turned Out to Be Enough

Grace shipped. The Arm-based data center CPU that Nvidia announced in April 2021 became a production part and then the CPU half of Nvidia's Grace Hopper and Grace Blackwell systems, in which an Arm processor and an Nvidia accelerator are packaged as a single computing unit and sold together. Nvidia designed those cores under the architecture license it had held before the deal and kept after it.

That outcome is the strongest available evidence against the industrial necessity of the acquisition. Everything Nvidia said it needed the CPU for, it built. What ownership would have added was control over what everyone else could build, which is precisely the thing the regulators objected to. The two propositions cannot easily be separated: the efficiency case and the foreclosure case were descriptions of the same integration, seen from the inside and the outside.

From Three Hundred Billion to Five Trillion

Nvidia ended calendar 2020 with a market capitalization of about $323 billion, close to eight times the price it had agreed to pay for Arm. By August 2026 the company was worth roughly $5.3 trillion, having become the most valuable listed company in the world; it closed at a record above $5 trillion in April 2026 and set a further record above $5.5 trillion in May.

The failed acquisition contributed nothing to that and cost very little. Against the run in the equity, the $1.35 billion write-off is a rounding error, and the seventeen months of executive attention consumed by the review coincided with the period in which Nvidia's data center business began compounding at a rate that made every other question secondary.

Selling the Last of the Arm Shares

There is a quiet coda. Nvidia had taken a small stake in Arm as a cornerstone investor in the 2023 IPO, and it wound that position down over the following two years. A regulatory filing disclosed in February 2026 showed Nvidia had sold its remaining 1.1 million Arm shares, worth about $140 million, taking its holding to zero.

The relationship survived the equity exit intact. Nvidia remains one of Arm's most visible architecture licensees, and Arm technology sits inside Nvidia's flagship data center platforms. The commercial arrangement that Nvidia said in 2020 was insufficient has turned out to be entirely sufficient, which is the single most important fact in any assessment of whether blocking the deal harmed anyone.

The parties disagreed at the time and the record has since resolved some of the disagreement, though not all of it.

The Case That the Regulators Read It Correctly

The strongest argument for the interventions is that the harm they described was structural rather than behavioral, and therefore not fixable by promises. Arm's value to the industry is that it is indifferent between customers. An owner that competes with those customers has an interest that is not indifferent, and the CMA's specific objection, that no monitoring trustee can reliably detect a research roadmap being tilted, is difficult to answer.

The independent analysis available before any regulator acted made the same point. SemiAnalysis argued in August 2020, weeks before the deal was even announced, that Nvidia's data center ambitions would push it to convert Arm from a low-margin neutral licensor into a vehicle for proprietary control, replacing the open ecosystem with a closed one over successive product generations. The mechanism described was gradual, commercially rational and effectively invisible from outside, which is exactly the profile that behavioral remedies handle worst.

Post-deal evidence supports the neutrality thesis in one important respect. Arm's revenue grew from $2.70 billion in the fiscal year in which the deal was abandoned to $4.92 billion in the year to March 2026, with data center royalties among the fastest-growing lines, in a period when its licensees have included every major cloud provider and several direct Nvidia competitors. That growth came from being the supplier all of them could trust, and it is at least consistent with the view that independence was worth more than integration.

The Case That an Efficient Deal Was Blocked

The counter-argument is that no regulator ever proved a case, and that the deal was killed by process rather than by evidence. The FTC never went to trial. The European Commission never issued a decision. The CMA never completed phase 2. The theories of harm were plausible predictions about future conduct, and predictions of that kind had been treated for decades as the weakest category of merger challenge because vertical integration usually lowers costs.

Nvidia's supporters can also point to what the ecosystem lost. Arm's licensing model funded incremental improvement, not the deep architectural investment that a company earning Nvidia's margins could have financed, and a faster Arm roadmap in the data center would have arrived earlier under Nvidia's ownership than it did under public-company discipline. That claim is untestable, which is precisely the problem with it.

The sharper version of the objection is procedural rather than substantive. A transaction that four authorities were still investigating, none of which had ruled, was defeated by the arithmetic of parallel timetables. If that is now the standard outcome for large cross-border technology deals, the practical effect is a de facto prohibition rule that no legislature enacted and no tribunal applied.

The Ledger, Party by Party

SoftBank is the unambiguous winner, and by a wide margin. It collected $1.25 billion for a deal that never closed, retained an asset it had been willing to sell at $40 billion, raised $4.87 billion in the IPO while keeping roughly 90% of the equity, and holds a position whose market value in August 2026 is many multiples of the price it had agreed to accept. That outcome was not foresight; Son tried to sell.

Nvidia is the second winner, which is the counter-intuitive part. It lost $1.35 billion, kept the license that mattered, avoided the integration risk and the customer conflict that ownership would have created, and compounded its own equity at a rate that made the Arm question irrelevant. The complainants achieved their objective and paid nothing for it. Arm's licensees preserved a neutral supplier and then bought shares in it.

Arm's own position is the most genuinely improved. It emerged with a new chief executive, a public listing, a Chinese joint venture moved off its own balance sheet and the strategic freedom to move up the value chain into complete compute subsystems, none of which would have been available to it inside Nvidia.

What Is Settled, and What Is Not

Two things are settled. Nvidia did not need to own Arm to execute its data center strategy, which the Grace, Grace Hopper and Grace Blackwell product lines demonstrate directly. And SoftBank was better off with the deal dead, which the difference between $40 billion and Arm's subsequent market value establishes beyond argument.

One thing is not settled, and probably cannot be. Whether Nvidia would have degraded Arm's neutrality is a claim about a world that does not exist. The regulators' case rested on incentive rather than conduct, and no evidence of actual foreclosure was ever produced because there was never an opportunity for any. Anyone asserting confidently that the FTC was right, or that it was wrong, is asserting something the record does not contain.

What the case does establish, and what makes it worth studying alongside deals that closed, is the mechanism of failure. Nvidia negotiated a fair price, structured the consideration intelligently for a seller with a liquidity need, pre-committed to remedies before anyone asked, and lost anyway, because a transaction requiring clearance in four jurisdictions is exposed to the slowest of them. The lesson is not that the price was wrong. It is that on deals of this shape, regulatory deliverability is a term of the deal, and it should be diligenced before the valuation is.

1. 1NVIDIA, "NVIDIA to Acquire Arm for $40 Billion, Creating World's Premier Computing Company for the Age of AI" (September 13, 2020).
2. 2Jensen Huang, "Letter From Jensen: Creating a Premier Company for the Age of AI", NVIDIA (September 2020).
3. 3TechCrunch, "Arm Holdings confirms SoftBank is buying the chip designer for £24.3B" (July 17, 2016).
4. 4Forbes, "SoftBank Announces $41 Billion Asset Sale To Cut Down Debt And Buy Back Shares" (March 23, 2020).
5. 5CNBC, "Nvidia's Arm takeover sparks concern in the UK" (September 14, 2020).
6. 6CNBC, "Qualcomm objects to Nvidia's $40 billion Arm acquisition" (February 12, 2021).
7. 7UK Competition and Markets Authority, NVIDIA / Arm merger inquiry case page.
8. 8UK Department for Digital, Culture, Media and Sport, "Digital Secretary asks CMA to carry out further investigation into NVIDIA's takeover of Arm" (November 16, 2021).
9. 9The Stack, "NVIDIA promised just 5 years of Arm independence: CMA" (2021).
10. 10Data Center Dynamics, "European Commission opens in-depth investigation into Nvidia's proposed Arm acquisition" (October 28, 2021).
11. 11Federal Trade Commission, "FTC Sues to Block $40 Billion Semiconductor Chip Merger" (December 2, 2021).
12. 12Cleary Antitrust Watch, "NVIDIA/Arm Transaction Collapse Signals Increased Scrutiny For Vertical Mergers" (February 2022).
13. 13Bloomberg, "Nvidia Is Said to Withdraw From Acquisition of SoftBank's Arm" (February 2022).
14. 14NVIDIA, "NVIDIA and SoftBank Group Announce Termination of NVIDIA's Acquisition of Arm Limited" (February 8, 2022).
15. 15SoftBank Group, "Termination of Agreement to Sell All Shares in Subsidiary Arm and Preparation for Arm Public Offering" (February 8, 2022).
16. 16Arm, "Arm Appoints Rene Haas as Chief Executive Officer" (February 8, 2022).
17. 17The Register, "Nvidia on the Arm termination, Q4 FY2022" (February 17, 2022).
18. 18NVIDIA, "NVIDIA Announces Financial Results for First Quarter Fiscal 2023" (May 2022).
19. 19NVIDIA, "NVIDIA Announces CPU for Giant AI and High Performance Computing Workloads" (April 12, 2021).
20. 20Arm Holdings Ltd, Form F-1 registration statement, SEC EDGAR (August 21, 2023).
21. 21CNBC, "SoftBank is in talks to buy Vision Fund's 25% stake in Arm" (August 13, 2023).
22. 22Arm, "Arm Announces Pricing of Initial Public Offering" (September 13, 2023).
23. 23CNBC, "Arm climbs 25% in Nasdaq debut after pricing IPO at $51 a share" (September 14, 2023).
24. 24Arm, "Arm delivers record-breaking quarter and full-year results" (May 6, 2026).
25. 25Bloomberg, "Nvidia Sells Off Stake in Arm, a Company It Once Tried to Buy" (February 17, 2026).
26. 26CNBC, "Nvidia stock closes at record, pushing market cap past $5 trillion" (April 24, 2026).
27. 27SemiAnalysis, "Jensen Huang's Vision For Data Center Dominance Will Destroy The Arm Ecosystem" (August 21, 2020).
28. 28Mobile World Live, "Arm regains control of China joint venture" (2022).