Amazon Web Services CEO Adam Selipsky speaks with Anthropic CEO and co-founder Dario Amodei during AWS re:Invent 2023, a conference hosted by Amazon Web Services, at The Venetian Las Vegas in Las Vegas on Nov. 28, 2023.
Noah Berger | Getty Images
Almost three years into a largely dormant IPO cycle, venture capitalists are in a tough spot.
The private market is dotted with richly valued artificial intelligence startups, including some that are described as generational companies. But venture firms in need of exits aren’t going to get relief from AI anytime soon.
That’s because, unlike prior tech booms, VCs aren’t at the center of this one. Rather, the biggest companies in the industry — Microsoft, Amazon, Alphabet and Nvidia — have been pouring in billions of dollars to fuel the growth of capital-intensive companies like OpenAI, Anthropic, Scale AI and CoreWeave.
With some of the most well-capitalized companies on the planet flinging open their wallets to fund the generative AI craze, the normal pressures to go public don’t apply. And even if they did, this batch of startups is nowhere near showing off the profitability metrics that public investors need to see before taking the plunge.
Tech giants have more than money. They’re also throwing in tangible benefits like cloud credits and business partnerships, packaging the types of incentives that VCs can’t match.
“The AI startups we talk to are having no problems fundraising at robust valuations,” Melissa Incera, an analyst at S&P Global Market Intelligence, told CNBC. “Many are still reporting having too much unsolicited investor interest at the moment.”
Add it all up and venture investors are maneuvering through a deep market distortion with no clear end in sight. U.S. VC exit value this year is on track to reach $98 billion, down 86% from 2021, according to an Aug. 29 report from PitchBook, while venture-backed IPOs are expected to be at their lowest since 2016. Traditional VCs are actively trying to play in AI, but they’re mostly investing higher up the so-called stack, putting money into nascent startups building applications that require far less capital than the infrastructure businesses powering generative AI.
So far in 2024, investors have pumped $26.8 billion into 498 generative AI deals, including from strategic investors, according to PitchBook. That continues a trend from 2023, when generative AI companies raised $25.9 billion for the full year, up more than 200% from 2022.
According to Forge Global, which tracks private market transactions, AI as a percentage of total fundraising jumped from 12% in 2023 to 27% so far this year. The average round for AI companies is 140% bigger this year compared to last, the data shows, while for non-AI companies the increase is only 10%.
Chip Hazard, co-founder of early-stage firm Flybridge Capital Partners, says investing dollars are shifting “up the stack” and that “enduring companies will be built at the application layer.”
That’s all going to take time to develop. In the meantime, startup investors continue to suffer from the fallout of the market turn that began in early 2022, when soaring inflation led the Federal Reserve to lift interest rates, pushing investors out of risky assets and into more conservative investments that finally offered yield.
Tech stocks have since bounced back, driven by Nvidia, whose chips are used in training most of the AI models, and other mega-cap stocks like Microsoft, Meta and Amazon. The Nasdaq hit a record in July before selling off a bit of late. But IPOs and pricey acquisitions have been few and far between, leaving venture firms with minimal returns for their limited partners.
“Managers are having a difficult time raising additional funds without delivering LP returns, especially because more liquid, lower-risk investments now have attractive yields thanks to high interest rates,” PitchBook wrote in its August report.
The one pure AI company that appears close to going public is Cerebras, a chipmaker founded in 2016 that’s backed by some traditional VCs including Benchmark and Foundation Capital. As a semiconductor company, Cerebras never reached the lofty valuations of the AI model developers and other infrastructure players, topping out at $4 billion in 2021, prior to the market’s downward tilt.
Cerebras said in late July that it had confidentially filed its IPO paperwork with the SEC. The company still hasn’t filed its public prospectus. A Cerebras spokesperson declined to comment.
When it comes to the foundational model companies, the astronomical valuations they quickly commanded put them in a very “different league,” outside of the realm of VCs, said Jeremiah Owyang, a partner at Blitzscaling Ventures.
It’s “very challenging for VCs to be promising any exits right now, given the market conditions,” Owyang said, adding that early-stage investors may not see returns for seven to 12 years on their newer bets. That’s for their companies that ultimately succeed.
Elbowing into big rounds
Firms like Menlo Ventures and Inovia Capital are taking another route in AI.
In January, Menlo disclosed that it was raising a so-called special purpose vehicle (SPV) — called Menlo Inflection AI Partners — as part of a $750 million funding round in Anthropic in a deal that valued the company at more than $18 billion. Since Anthropic’s launch in 2021, Amazon has been the company’s principal backer as it tries to keep pace with Microsoft, which has poured billions of dollars into OpenAI and is reportedly part of an upcoming funding round that will value the ChatGPT creator at over $100 billion.
Menlo had previously invested in Anthropic in 2023 at a valuation of about $4.1 billion. To put in more money at a much higher price, Menlo had to go outside of its main $1.35 billion fund that closed last year. In raising an SPV, a venture firm typically asks for LPs to put money into a separate fund dedicated to a specific investment, rather than a portfolio of companies. Menlo filed to $500 million for the SPV.
In July, rival startup Cohere, which focuses on generative AI for enterprises, announced a $500 million funding round from investors including AMD, Salesforce, Oracle and Nvidia that valued the company at $5.5 billion, more than doubling its valuation from last year.
Cohere confirmed to CNBC that part of the financing, as well as some of its previous fundraising, came through an SPV. Inovia, based in Montreal, organized the latest SPV, and Shopify CEO Tobias Lutke was one of the participants.
Representatives from Menlo and Inovia didn’t respond to requests for comment.
Some investment banks have also put together SPVs to allow multiple investors to pool capital into a hot company. JPMorgan Chase told CNBC that clients “have been able to access several leading AI investments” through the bank’s Morgan Private Venture unit.
Still, for investors to get a return there has to be an IPO at some point, as the regulatory environment makes it virtually impossible for big tech companies to orchestrate significant acquisitions. And companies like Microsoft, Alphabet, Amazon and Nvidia can be plenty patient with their investments — they have a combined $280 billion in cash and marketable securities on their balance sheets.
IPO pipeline will ‘continue to build’
The other potential path for liquidity is the secondary market, which involves selling shares to another investor.
Elon Musk’s SpaceX, which reportedly valued itself at over $200 billion in a recent employee tender offer, has enabled investor shares through secondary transactions. That may be what’s eventually in store for some investors in xAI, Musk’s 18-month-old AI startup, which is already valued at $24 billion after raising a $6 billion round in May.
But SpaceX is an outlier. For the most part, secondary transactions are viewed as a way for founders and early investors to cash out a portion of their stock in a high-valued company, not a way for VCs to generate returns. For that they need IPOs.
SpaceX’s Polaris Dawn Falcon 9 rocket sits on Launch Complex 39A of NASA’s Kennedy Space Center on August 26, 2024 in Cape Canaveral, Florida.
Joe Raedle | Getty Images
Michael Harris, global head of capital markets at the New York Stock Exchange, told CNBC recently that NYSE is in dialogue with “a number of AI-focused companies” and said that, “as the industry evolves we’d expect that pipeline to continue to build.”
A select few AI companies have hit the public market this year. Astera Labs, which sells data center connectivity to cloud and AI infrastructure companies, debuted on the Nasdaq in March. The company is valued at about $6.5 billion, down from $9.5 billion after its first day of trading.
Tempus AI, a health-care diagnostics company backed by Google, went public in June. The stock is up around 50% from its debut, valuing the company at $8.6 billion.
The IPO floodgates never opened, though, and high-profile AI companies aren’t even talking about going public.
“Unless there is a dramatic shift in market sentiment, I would be hard-pressed to see why these AI startups would put themselves in the public spotlight when they can keep growing privately at such favorable terms,” said S&P’s Incera. Going public “would only amp up pressure to show returns or reduce spending, which for a lot of them is not a feasible ask at this point in the maturity curve,” she said.
Most venture investors are bullish on the potential for generative AI to eventually create big returns at the application layer. It’s happened in every other notable tech cycle. Amazon, Google and Facebook were all web applications built on top of internet infrastructure. Uber, Airbnb and Snap were a few of the many valuable apps built on top of smartphone platforms.
John-David Lovelock, an analyst at Gartner and a 35-year veteran of the IT industry, sees a big opportunity for generative AI in the enterprise. Yet, in 2024, only 1% of the trillion dollars spent on software will be from businesses spending on generative AI products, he said.
“There is money being spent on certain GenAI tools and the few applications that exist,” Lovelock said. “However, broad-scale rollout of GenAI within the broad enterprise software catalogue of products has not yet occurred.”
At the Meta Connect developer conference, Mark Zuckerberg, head of the Facebook group Meta, shows the prototype of computer glasses that can display digital objects in transparent lenses.
Andrej Sokolow | Picture Alliance | Getty Images
Mark Zuckerberg on Wednesday unveiled the $799 Meta Ray-Ban Display glasses, the social media company’s first consumer-ready smart glasses with a built-in display.
The glasses, which costs $799, contain a small digital display that can be controlled via hand gestures through a wristband powered by neural technology, confirming a CNBC report in August. A promotional video of the new smart glasses appeared on Meta’s YouTube page on Monday but was later removed.
Tune in Thursday at 11:00 a.m. ET: Meta Chief Product Officer Chris Cox joins CNBC TV to discuss with Julia Boorstin the highlights of Meta’s annual Connect event, live from the company’s HQ in Menlo Park CA.
The new smart glasses are a bridge between the company’s audio-only Ray-Ban Meta smart glasses and the experimental Orion augmented reality glasses that the company revealed at last year’s Connect event. Orion can overlay 3D visuals over a person’s real-world field of view with the help of a wireless computing puck, but the glasses are expensive to make and not yet available to consumers.
The Meta Ray-Ban Display glasses come with the Meta Neural Band, an EMG wristband that allows users to control the device using hand gestures.
“These are glasses with the classic style that you’d expect from Ray-Ban, but they’re the first AI glasses with a high resolution display and a fully weighted Meta neural band,” Zuckerberg said.
With the new glasses, people can do tasks like watch videos through the display or see and respond to text messages, Zuckerberg said. The display doesn’t block a person’s view, and it disappears when not being used, he said.
The glasses go on sale in the U.S. on Sept. 30.
During a demo, Zuckerberg repeatedly attempted to call Meta tech chief Andrew Bosworth unsuccessfully.
“This is uh — you know, it happens,” Zuckerberg said.
Meta has been developing its smart glasses with eyewear giant EssilorLuxottica since 2019, and last year renewed a long-term partnership agreement to continue making the products.
The company on Wednesday also debuted the Oakley Meta Vanguard smart glasses, intended for athletes who participate in high-intensity sports like snowboarding and mountain biking. The Oakley-branded glasses will cost $499 when they launch on Oct. 21, making it $100 more expensive than the Oakley Meta HSTN glasses that went on sale in June.
The Oakley Meta Vanguard smart glasses have a sportier look than the Oakley Meta HSTN glasses thanks to a wraparound design that extends its colorful lenses around a person’s temples. Unlike the Oakley Meta HSTN glasses, the new model contains a button on the underside of its frames so that athletes who wear helmets can more easily capture photos and videos.
The new sports-centric smart glasses have up to nine hours of battery life, can capture 3K video and contain speakers that are louder than their predecessors. The glasses can connect with Garmin-branded fitness watches to track certain stats like their heart rates using the Meta AI assistant. Preorders start today.
Meta also debuted the Ray-Ban Meta (Gen 2), the latest version of the company’s original smart glasses. The Ray-Ban Meta (Gen 2) costs $379, up from $299 for the version released in 2023. The Ray-Ban Meta (Gen 2) has double the battery life of the previous model, lasting 8 hours on a single charge, and a more powerful camera that can capture 3K Ultra HD video. The new glasses go on sale today.
Zuckerberg also announced Horizon TV, pitching it as a way to watch television shows, sporting events and movies using the company’s Quest VR headsets. Some of Meta’s partners who will be contributing content to the app include Disney and Universal Pictures, Zuckerberg said.
Thomas Fuller | SOPA Images | Lightrocket | Getty Images
Cybersecurity company Netskope is eying a $7.3 billion valuation after pricing shares at $19 for its upcoming IPO, at the top end of its expected range.
Netskope will start trading on Thursday on the Nasdaq under the ticker symbol “NTSK.” The share sale raised $908.2 million.
Earlier this week, Netskope lifted its expected pricing range to between $17 and $19 a share, up from an original range of $15 to $17. The company revealed plans to go public last month.
Netskope’s offering comes amid a hot period for IPO activity after a years-long lull spurred by step inflation and soaring interest rates. The long-overdue resurgence has fueled optimism on Wall Street and in a venture capital industry eager for return on investment.
Ticket reseller StubHub slid 6% it its first day of trading Wednesday, but a lackluster start may not be reason for concern. CoreWeave went public in March and closed flat in its first day, with shares going on to triple.
Swedish buy now, pay later firm Klarna jumped 15% in its debut this month. Peter Thiel-backed cryptocurrency exchangeBullish, design software company Figma and stablecoin issuer Circle have also jumped since their recent market debuts.
Read more CNBC tech news
The cybersecurity sector is also undergoing a busy stretch for dealmaking fueled by ongoing artificial intelligence advancements and a shifting threat landscape.
Santa Clara, California-based Netskope was founded in 2012 and is led by co-founder and CEO Sanjay Beri. At the end of July, the company said it had 2,910 employees and 4,317 customers across 90 countries.
Annual recurring revenues rose 33% to $707 million at the end of July and revenues reached $328 million for the six months ended July 31. The company also reported a net loss of $170 million during that period.
Some of Netskope’s significant backers include Accel, Iconiq and Lightspeed Venture Partners.
Nscale, the UK-headquartered AI infrastructure provider.
Courtesy: Nscale
Two years ago, Nscale was a brand new startup in the U.K. that had yet to raise any outside funding or officially announce its existence.
Last year the London-based company came out of stealth, and in December announced that it had raised its Series A fundraising, totaling $155 million.
Now, Nscale finds itself at the center of the action in the hottest market on the planet: artificial intelligence. And it has close to $700 million in fresh capital from Nvidia, the world’s most valuable company.
In press releases on Tuesday, Nscale was named as an AI infrastructure partner for Nvidia, Microsoft and OpenAI, as the companies expand their buildouts in the U.K. Nscale then said it signed a five-year $6.2 billion agreement with Microsoft and Aker to develop “hyperscale AI infrastructure” in Europe, specifically Norway, where Aker is headquartered.
OpenAI made prior headlines with Nscale, announcing plans in July for a data center in Norway for a Stargate-branded AI data center. Nscale agreed to commit $1 billion for the project, with the goal of racking up 100,000 Nvidia graphics processing units (GPUs) at the site before 2027.
It’s a remarkably quick rise for a company that wasn’t even around when OpenAI kicked off the generative AI boom with the launch of ChatGPT in late 2022. At that time, what’s now Nscale was part of Arkon Energy, which was established a year earlier to provide infrastructure for cryptocurrency mining. Nscale was spun out to address soaring demand for data centers capable of handling AI workloads.
Read more CNBC tech news
Like CoreWeave, which went public this year and now sports a market cap of $58 billion, Nscale is combining data center space, power and lots of GPUs with its own software in order to an provide end-to-end service for AI infrastructure.
CoreWeave, which supplies infrastructure to Microsoft, Google, Nvidia and OpenAI, also has roots in crypto. Founded in 2017, the company built up its initial fleet of Nvidia GPUs for ethereum mining before pivoting to AI.
Nscale didn’t respond to a request for comment following this week’s announcements, but CEO Josh Payne, who previously founded Arkon, told CNBC in late July that the company was targeting two big problems in Europe. One is a lack of sufficient computing capacity and the other is a “very fragmented market.”
“What the continent needs is large AI infrastructure projects deploying compute [power],” Payne said, after the announcement with OpenAI for the Norway buildout. “The ecosystem can consume from the project to build AI products, to generate productivity growth and economic benefit.”
Payne wrote in a LinkedIn post on Wednesday that the agreement with Microsoft and Aker is a “huge win for European-owned AI infrastructure.”
Europe has been pushing the concept of “sovereign AI,” requiring data centers and AI workloads to be located and processed on European soil. Nscale has quickly emerged as an important player in the U.K.’s bid to evolve into a global leader in AI. In January, Britain laid out an AI “action plan,” promising to reduce bureaucracy to help its domestic AI sector thrive.
While Nscale is addressing the European market, many of its early partners are big U.S. AI vendors. They timed their announcements on Tuesday to President Donald Trump’s state visit to the U.K.
On Wednesday, Trump visited Windsor Castle and met with King Charles, Queen Camilla and other members of the royal family. His trip comes at a contentious moment for U.K. Prime Minister Keir Starmer, who is under pressure to bring stability to the country after the exit of Deputy Prime Minister Angela Rayner over a house tax scandal and a major cabinet reshuffle.
Microsoft headlined the U.K. announcements, committing $15.5 billion of new investment to computing equipment. The software giant said it plans to work with Nscale to construct what will become the U.K.’s largest supercomputer in Loughton, a suburban town in the English county of Essex.
The site will initially house 23,040 Nvidia Blackwell GPUs to be delivered in the first quarter of 2027. When it goes live, it will generate 50 megawatts of AI capacity, scalable to 90 megawatts, according to a statement from Nscale.
“No one can make that kind of capital investment unless they’ve got somebody already committed to spend the money once the work is complete, and that’s the role we’re playing,” Microsoft President Brad Smithsaid Tuesday, adding the deal represents a major vote of confidence in Nscale.
OpenAI said it would launch a U.K. version of Stargate through a partnership with Nscale and Nvidia. OpenAI will deploy 8,000 GPUs in the project’s first phase early next year, with the option to expand capacity to approximately 31,000 GPUs over time.
Stargate U.K. will operate across a number of sites in the country — one of the early ones being Cobalt Park, an industrial state in the Northern English city Newcastle. Stargate was initially spawned in the U.S. in January as part of President Trump’s effort to push investments in AI infrastructure.
Nvidia CEO Jensen Huang attends the “Winning the AI Race” Summit in Washington D.C., U.S., July 23, 2025.
Kent Nishimura | Reuters
Nvidia’s announcement on Tuesday included an investment of up to £11 billion ($15 billion) with Nscale and CoreWeave to boost U.K. AI infrastructure.
Nvidia CEO Jensen Huang separately revealed on Wednesday that the chipmaker had made a £500 million ($683 million) equity investment into Nscale.
“We convinced ourselves that Nscale could be a national champion for AI infrastructure in the U.K.,” Huang told journalists at a press conference in London.
Nick Patience, AI practice lead at the Futurum Group, told CNBC that Nscale is “a key part of Nvidia’s push in the U.K. market and an acknowledgment by the government that it has to do something to get the AI infrastructure built here, which has been a long slog.”
Rapid growth
After exiting stealth in May of last year, Nscale’s first public announcement came two months later, when the company partnered with UAE’s Open Innovation AI to deploy 30,000 GPUs. Around the same time, Nscale said it was acquiring Kontena, which was founded in 2018 and specialized in high-performance computing data centers.
The next month, Nscale announced an agreement with Asian telecom company Singtel to offer a “GPU-as-a-Service (GPUaaS),” and serve customers in Europe and Southeast Asia. Initially, Nscale’s infrastructure relied on GPUs from Advanced Micro Devices. Today, the startup promotes various offerings from market leader Nvidia.
Nscale’s big financing landed in December, when the company said it raised $155 million in a round led by Sandton Capital Partners, with participation from Kestrel0x1, Blue Sky Capital Managers and Florence Capital.
Sandton co-founder Rael Nurick said in the press release that with its “unique vertically integrated approach, Nscale is building the hyperscale AI platform to power AI at scale.”
Nscale said at the time that it had grown its AI data center pipeline to 1.3 gigawatts from 300 megawatts the prior year to and that it was aiming to have 350,000 GPUs running by the end of 2027.
By comparison, CoreWeave said at a banking conference last week that its portfolio consists of “about 2.2 gigawatts of capacity that’s coming online.” The company said in its IPO prospectus in March that its 32 data centers were running 250,000 GPUs.
It’s been a whirlwind few years for Payne, Nscale’s founder. While he was serving as executive chairman of Arkon, he was also operating chief at Australia’s Battery Future Acquisition Corp., a blank check company that says it’s “targeting critical battery minerals and related supply chains.”
He’s got a lot of work in front of him.
Building out AI data centers with costly GPUs is a capital intensive process that’s historically required a hefty amount of debt. CoreWeave had raised a total of $12.4 billion in debt through the end of 2024, in addition to well over $1 billion in equity financing before its IPO. It announced a $1.5 billion bond sale in July after a $2 billion debt offering in May.
Nscale was trying to raise $1.8 billion earlier this year through a private credit deal led by bankers at Goldman Sachs, according to Bloomberg.
In the December video tied to Nscale’s equity fundraising, Payne called it “one of the largest Series As raised in U.K., European history.” He said the company would use the cash to deploy up to another 4,000 GPUs in its data center in Norway and to develop up to 180 megawatts of capacity in the company’s portfolio.
The aim, Payne said, was to deploy 50,000 GPUs by the end of 2025 and 150,000 by the end of next year.
“The key challenges that we see in the market is the significant increase in density at the GPU level,” he said. “This funding allows us to scale up materially” he said, and to become “one of the largest players in Europe.”