2016 to 2026
2016
Smartphone shipments grew about three percent in 2016, against roughly ten percent in 2015, and the first annual decline came in 2017. A company whose main market doubles every few years can fund its growth from that market. A slower market leaves large technology companies choosing between buying their way into adjacent ones and accepting lower growth. The first two acquisitions below were made in that setting.
Acquisition · $32bn · announced July 2016
SoftBank / ArmRead the full note →Arm designs the instruction set that processors are built around and licenses it. It manufactures nothing. Apple, Qualcomm, Samsung and others take the design, build a chip around it and pay Arm a royalty on each one sold, a few cents a chip, on more than fifteen billion chips a year.
SoftBank paid about $32 billion in cash, a premium of roughly 43 percent, for a company whose royalties came mostly from a market that was flattening. Because Arm owns no factories it carries none of the capital spending of the rest of the industry, and because it makes no chips it does not compete with its licensees.
The price was a bet on the licensing structure more than on mobile. Arm's designs were later used in server processors from Amazon and Nvidia, and SoftBank listed Arm again in 2023 at about $54.5 billion.
2017
Acquisition · $15.3bn · announced March 2017
Mobileye made the camera systems and software that let a car recognize a lane, a pedestrian or a vehicle ahead. It sold to carmakers, who designed it into models years before they reached showrooms, which made its revenue predictable once a design win was secured.
Intel paid about $15.3 billion against Mobileye revenue of around $358 million the previous year, more than forty times. A buyer paying that multiple is assuming either that revenue grows by a large multiple or that the asset has value beyond its sales. Intel said publicly that it saw the car as a future computing platform, and it had missed mobile almost entirely.
This is the clearest case on the page of a price resting on growth that had not yet occurred, and it is the objection to the thesis that the closing section addresses.
Carve-out · ~$18bn · signed September 2017
Toshiba sold its memory chip business to a consortium led by Bain Capital after losses in its nuclear division left it short of cash. The business made NAND flash, the storage in phones, laptops and data centers.
The sale took close to a year to complete because several governments took an interest in who would own it.
Government interest in who owns a chip business, which shows up here before it does in the larger cases, became routine over the decade.
2018
Prohibited · ~$117bn · March 2018
Broadcom / QualcommRead the full note →Broadcom pursued Qualcomm without its board's agreement and raised its offer to around $117 billion. Qualcomm holds a large body of patents covering how mobile devices communicate with networks and was heavily involved in writing the 5G standards.
A presidential order prohibited the transaction in March 2018 after a review by the Committee on Foreign Investment in the United States. The committee's concern was that Broadcom's practice of cutting research spending at companies it buys would weaken American participation in those standards.
Approval risk shifted from a market share question a bidder could model to a forecast of national capability it could not.
Abandoned · $44bn · July 2018
Qualcomm agreed to buy NXP Semiconductors in 2016 and spent close to two years collecting approvals. All arrived except Chinese approval, and in July 2018 Qualcomm let the agreement lapse and paid NXP a $2 billion termination fee.
A fee of that size is agreed in advance because approval is uncertain. It compensates the seller for the time spent waiting and for being publicly in play.
The second abandoned deal of the year also ended in a regulator's office. Both failures were about positions in the supply chain.
Acquisition · $34bn · announced October 2018
IBM / Red HatRead the full note →Red Hat's main product is a version of Linux, which is free software. What it sold was a version somebody would stand behind: tested, patched, supported and certified to work with the other systems a bank or a hospital runs.
That is a subscription business, and its strength comes from what a customer would have to do to leave. An institution that has certified several hundred applications against one distribution would have to test all of them again. IBM paid $190 a share in cash, about $34 billion, a premium above 60 percent, for software that enterprises use across more than one cloud provider.
The case for the price is the customer's cost of leaving. The risk is whether software valued for working equally well with every cloud can be owned by a company that sells one.
2019
Acquisition · $6.9bn · announced March 2019
Nvidia / MellanoxRead the full note →Mellanox made high speed networking equipment, mainly a technology called InfiniBand, designed for connecting computers inside a single facility. Its customers were mostly scientific computing centers running weather models, physics simulations and genomics work.
A large job split across many machines requires them to stop periodically and exchange results. If the connection is slow the processors spend their time waiting, so networking sets how much of a facility's equipment is in use. Nvidia paid $6.9 billion in cash, its largest acquisition to that point.
The purchase is explained by what Nvidia's customers were already doing in 2019. The AI training demand that followed made the capability far more valuable, and the announcement does not claim to have forecast it.
Acquisition · $15.7bn · announced June 2019
Tableau made software for turning company data into charts that non-specialists could build themselves. Salesforce paid about $15.7 billion in its own shares.
Analysis software tends to be sticky because an organization that has built hundreds of reports finance and sales rely on weekly would have to rebuild all of them and retrain its staff to move to a competitor. The switching cost sits in the work already done.
Another instance of the cost-of-leaving test, here applied to software that is used by employees and not run by engineers.
2020
Central banks cut interest rates to near zero in response to the pandemic and held them there. Cheap borrowing raises what buyers can afford and raises the price of assets whose value depends on earnings far in the future. Most of the transactions in the next two years were negotiated under those conditions.
Cheap money is a competing explanation for the high prices of 2020 and 2021, and the closing section treats it as an objection.
Acquisition · $35bn at announcement · October 2020
AMD / XilinxRead the full note →Xilinx made field programmable gate arrays, chips that can be reconfigured after manufacture. They are slower than a chip built for one purpose and more flexible, which suits applications where the task changes, such as telecommunications equipment.
AMD paid entirely in its own shares at a fixed ratio of 1.7234 AMD shares for each Xilinx share, giving Xilinx holders about 26 percent of the combined company. Because the ratio was fixed, the value of the consideration moved. The deal was described as about $35 billion at announcement and closed in February 2022 at a figure widely reported near $49 billion.
The announced value of an all-stock deal is the buyer's share price on one day multiplied by a ratio. The ratio and the ownership split are the terms, and comparisons that use the announced dollar value mislead.
2021
Global deal volumes reached their highest level on record in 2021, helped by cheap debt, high equity prices and a large number of blank check vehicles raised to buy companies. Several transactions agreed that year were renegotiated or abandoned once conditions changed.
Acquisition · $19.7bn · announced April 2021
Nuance sold speech recognition software, with its strongest position in hospitals, where it turns a clinician's dictation into a patient record. Microsoft paid $56 a share in cash, about $19.7 billion including debt.
The attraction was less the software than the access. Health systems are slow and cautious about who they allow near patient data, for sound regulatory reasons, and a supplier that is already installed and approved holds a position that is difficult to buy any other way.
The asset was an approved position inside a regulated customer, which is a switching cost created by the customer's own rules.
2022
Interest rates rose quickly through 2022. The Nasdaq composite fell by roughly a third over the year, debt became more expensive and the market for initial public offerings effectively shut. Companies that had expected to list stayed private, which is where a growing share of large technology businesses has remained.
Abandoned · $40bn · February 2022
Nvidia / ArmRead the full note →Nvidia agreed in September 2020 to buy Arm from SoftBank for up to $40 billion in cash and stock. Seventeen months later the parties abandoned it, and SoftBank kept a $1.25 billion payment Nvidia had made in advance.
Arm's licensees include most of Nvidia's competitors, and they license from Arm partly because Arm does not build chips. Several objected to regulators in the United States, the United Kingdom and the European Union, and the US Federal Trade Commission sued to block the deal. SoftBank listed Arm on Nasdaq in September 2023 instead, at a valuation above what Nvidia had offered.
Where part of an asset's value comes from its independence, a sale to an interested party reduces what is being bought.
Acquisition · $61bn · announced May 2022
Broadcom / VMwareRead the full note →VMware's core product lets one physical server run as several separate machines. It sits underneath a large amount of corporate computing and is difficult to remove without a long migration project.
Broadcom has bought established infrastructure software repeatedly, including CA Technologies and Symantec's enterprise business. It reduces research spending and smaller product lines, concentrates on the largest customers and raises prices, so its return does not depend on the acquired business growing. The VMware transaction was announced as rates rose, took eighteen months of review and completed in November 2023. Customers and cloud providers have since complained about large price increases.
The price is a purchase of customers who cannot leave cheaply. Whether that approach carries a delayed cost with customers is not yet established.
2023
ChatGPT was released in late 2022 and reached wide use through 2023. The consequence that mattered most to the companies on this page was the computing it required. Training a large model uses thousands of processors running for weeks, and demand for that capacity grew faster than the infrastructure to supply it.
Technology deal volumes in 2023 were the weakest in more than a decade. High rates made debt expensive, valuations agreed in 2021 had not yet adjusted, and buyers and sellers disagreed about price. Periods like this are common after a sharp change in the cost of money and usually end when one side gives way.
Acquisition · $28bn · announced September 2023
Cisco / SplunkRead the full note →Software records what it does, and a large organization produces enormous volumes of these records, called logs. Splunk collected them and made them searchable. Security teams pay for that because the evidence of an intrusion is scattered across thousands of systems, and engineering teams pay because modern applications are assembled from many services and diagnosing a failure means reading the record.
Cisco paid $157 a share in cash, about $28 billion, a premium of roughly 31 percent and its largest acquisition. It was announced in a year when very few large transactions were, and paid for in cash when cash was expensive.
Splunk is paid for from two budgets. The open question is Cisco's ability to sell and run subscription software after decades of selling hardware.
2024
Capital spending by the largest cloud providers rose sharply from 2024, and a growing share of it went to data center construction, electrical capacity and cooling and not to software development. Facilities of this kind are measured in megawatts of power draw, and in several markets the limiting factor became how long a utility took to provide a grid connection.
This is the third part of the thesis. The scarce input moved from ideas and processors toward electricity.
Acquisition · ~$1.46bn · announced March 2024
Spirent made equipment for testing networks, which generates realistic traffic so an operator can find out how a network behaves under load before customers do.
It is an ordinary industrial business. Its customers are equipment makers and network operators, it sells into long product development cycles and its position rests on having been validated against standards that change slowly. Keysight prevailed over a competing offer from Viavi Solutions and completed the purchase in October 2025, nineteen months after announcing it.
A validated position in a slow-changing standard is another form of switching cost, and a modest one, which is consistent with the modest price.
2025
Investment · $300m at a $5bn valuation · January 2025
DDN / BlackstoneRead the full note →DataDirect Networks sells storage systems to organizations running very large computing jobs. It was founded in 1998 and spent most of its history selling to national laboratories and research institutions.
A facility full of expensive processors is productive only while those processors have data to work on, so the customer is buying utilization of equipment it has already paid for. Blackstone Tactical Opportunities invested $300 million at a $5 billion valuation, and DDN said the money was for getting data to processors faster inside enterprise AI clusters.
A $5 billion valuation on a private company is an expectation. Whether DDN reaches the revenue it has described is what to watch.
Acquisition · $32bn announced · March 2025
Google / WizRead the full note →Wiz sells software that scans a company's cloud infrastructure and identifies which combinations of settings create a real risk. Earlier tools produced long lists of warnings that were mostly ignored, and Wiz produced short lists and could be returning results within hours.
Wiz ended discussions with Google in July 2024 at a reported $23 billion and stated an intention to go public. Eight months later it agreed to sell to the same buyer for $32 billion.
Little about the business changed in that time. What Wiz gained was a credible alternative, and a seller with somewhere else to go negotiates from a stronger position.
Acquisition · $31 per share · announced December 2025
Confluent sells a commercial version of Apache Kafka, software that moves data between systems continuously and not in overnight batches. IBM agreed to pay $31 a share in cash, an enterprise value of roughly $11 billion.
IBM's public case concerned AI systems that take actions and not only answer questions. Software of that kind needs current information, so the difference between data that is up to date and data that is a day old becomes operationally significant.
Much of the 2025 activity on these balance sheets was financing and not acquisition. Datadog issued convertible notes, Seagate issued exchangeable notes and LendingTree arranged a term loan. These transactions rarely make headlines and they paid for the spending described above.
A convertible note is a loan that the holder can exchange for shares on agreed terms. The borrower accepts possible dilution in return for a lower interest rate, and the lender accepts the lower rate in return for taking part if the shares rise. It suits companies whose shares are highly valued and whose cash flows are not yet large.
2026
Google completed the Wiz acquisition on 11 March 2026. Alphabet's quarterly filing puts the purchase price at approximately $29.5 billion after adjustments, against the $32 billion described at announcement a year earlier. IBM completed Confluent on 17 March 2026.
The gap between an announced figure and a final one is routine and is reported much less prominently than the announcement. Announced values are a starting point and not a record of what changed hands.
Hard drives were widely expected to be displaced by flash memory, which is faster and has no moving parts. Demand for very large and cheap capacity, driven by the volume of data retained for AI training, improved the position of the remaining manufacturers, and Seagate raised money on terms more often associated with growth companies.
Storage is a cyclical industry and current conditions are unusually favorable, so the entry describes a change in conditions and not a permanent reversal.
As of October 2026 the constraint most often described by the companies on this page is electrical capacity and the time required to obtain it, and not ideas or access to processors. Whether that remains true depends on how quickly utilities and regulators respond, which is not a technology question.
The band down the page is a hand-drawn outline of deal activity and not a data series. Values are announced values unless an entry says final.
