Coding Revenue and Compute Shortages Are Extending the AI Boom

Alex Sacerdote’s investment framework starts with three questions: where a technology sits on its adoption curve, whether a company has a durable competitive advantage inside that curve, and whether the market is underestimating its long-term earnings power. His AI thesis is an investor’s thesis, not an audit: AI is the largest technology adoption curve Whale Rock has studied, and the usual “S-curve” language may understate what is happening.

Enterprise AI, by Sacerdote’s estimate, is “less than 1% penetrated.” He distinguishes casual usage from deep workflow usage. Hundreds of millions of people may be using AI as “a search engine on steroids,” but that is not the same as enterprises wiring AI into tools, building skills, and deploying bots that perform work. He cites Sundar Pichai’s estimate that only about 10 basis points of global knowledge workers are using AI in that deeper sense.

That is why he calls it not an S-curve but an “L-curve” or “backwards L-curve”: a long horizontal period, followed by a line that appears to go straight up. The enterprise market is still at the tinkerer stage. In Sacerdote’s view, it then moves to early adopters and early mainstream users, with penetration potentially moving from 10 basis points to 2% or 3%, then 5%, then 15% over the next four years. He says a “light switch” has gone off inside the enterprise: executives now realize they need to move quickly.

The analogy is the late-1990s internet. In 1998, a company knew it needed a website, but building one was still difficult. AI, in Sacerdote’s telling, is at a similar institutional moment, except the building blocks are coming together faster.

The vertical adoption claim feeds directly into the infrastructure claim. Even before meaningful AI usage reaches 1% of knowledge work, Sacerdote says, “there’s not enough compute in the world.” Anthropic, he says, has only about half the compute it currently needs. He also cites Marc Andreessen’s view that one thing Andreessen is sure of over the next four years is that there will not be enough compute.

The point is not simply that AI demand is growing. It is that tiny penetration, visibly improving products, enterprise urgency, and already-constrained infrastructure make the early numbers misleading. If the adoption base is still this small while hardware supply is already short, Sacerdote sees the next several years as a compounding demand problem rather than a normal growth cycle.

Alex Sacerdote starts the Anthropic thesis with the stack that emerged after ChatGPT’s launch in November 2022. Whale Rock immediately did a “massive deep dive” with its 10-person team. Their first conclusion was that a new compute paradigm creates a new stack, and therefore new winners and losers.

Sacerdote describes the AI stack as power and chips at the bottom, then cloud, then foundational models, then applications. In early 2023, Whale Rock wanted exposure first to chips and infrastructure. Demand would reach those layers first, and the winners were more identifiable. Whoever won at the model or application layer would still need enormous compute.

The model layer was less clear. In April 2023, Whale Rock laid out multiple possibilities: winner-take-all, commodity because of open source, race to zero, or an oligopoly of three or four leading players. Over the next several years, Sacerdote says, the field began to resolve. Most startups “fell away and died.” Amazon, in his assessment, “never really showed up.” Meta came in strong but later faltered and had to reboot. Google’s Gemini could not be counted out, and Google remained one of Whale Rock’s largest positions. OpenAI had won the consumer market. Anthropic emerged as the dark horse, focused more purely on enterprise.

That led Whale Rock toward a three-horse-race thesis: OpenAI, Anthropic, and Google. Sacerdote compares it to cloud computing, where three large providers underpin the SaaS and cloud world.

The major remaining risk was open source, including models from China. Sacerdote argues that the frontier models’ token quality remains meaningfully superior. If a model is 80% of the way to the benchmark leader, moving from 80 to 85 can be a “huge unlock.” Open-source competitors can come close, but without comparable compute they struggle to leapfrog and maintain the frontier. Scaling laws, feedback loops, and other model-improvement mechanisms gave Whale Rock confidence that the leading labs still had runway.

But the decisive change was code.

Whale Rock knew AI could be large before 2025 and made large investments based on expected training demand, but Sacerdote says the firm remained skeptical about how much revenue would arrive and whether models could truly replace labor. Early versions were useful but not yet convincing enough for many corporates. They were not clearly agentic.

Coding tools changed that. The first generation, such as Microsoft Copilot at roughly $20 a month, could improve the “grammar” of coding, find bugs, or write a block of code — something like a paragraph. Then Claude Code and other coding products moved toward agentic behavior. Sacerdote says the market “just exploded.”

One data point mattered: Whale Rock heard that people using Anthropic tools unfettered were spending about $100 a day on tokens. Annualized, Sacerdote says, that becomes $20,000 to $30,000 per user. If there are roughly 20 million coders globally, coding alone could represent a half-trillion-dollar market. That estimate, in his framing, was based on technology that was only seven, eight, or nine months old.

Anthropic’s reported growth reinforced the point. Sacerdote describes numbers “like nothing we’d ever seen before,” saying the company had gone from “100 to a billion” and was “on the way to nine,” with Patrick O'Shaughnessy interjecting, “Yeah, one to nine.” Whale Rock invested at what Sacerdote describes as a $180 billion valuation in August 2025, at a time when he says “nobody had any idea what 2026 could be.”

The second coding unlock came when Claude Code became almost completely agentic. Sacerdote points to Andrej Karpathy and Linus Torvalds as sophisticated coders who had changed their views. The specific coding split he describes is Karpathy’s: last year, Sacerdote says, Karpathy’s tools could write about 20% of the code while 80% was handwritten; with the latest model, that flipped, and Karpathy said he had not written a line of code except in English. Sacerdote also emphasizes the additional market unlocked by people who never knew how to code at all.

In this model, coding strength is not a generic feature that all model providers quickly share. Foundational models have meaningful differentiation: different training methods, different strengths, and different product ecosystems. Routers that switch between models can make the layer appear commoditized, but Sacerdote says the differences matter. Anthropic is strong in code and in areas such as private equity and finance. Google is strong at ingesting PDFs. The model layer, in his view, contains “critical IP.”

Anthropic’s advantage is also not only the API. Sacerdote says the company is building a wider ecosystem: SDKs, Claude for Work, orchestration layers, tools, and what he calls a “harness” around the model. He compares this to AWS in 2013. Skeptics then saw AWS as commodity servers in a warehouse. Amazon, in his telling, saw a new way of computing and built products that created lock-in over time.

Alex Sacerdote’s framework depends on an uncomfortable premise for many investors: in the right part of a technology adoption curve, the future can be more predictable than it looks.

When a company is positioned correctly on a major S-curve and has a strong business model, unit growth becomes exponential and earnings can grow even faster. The result is underappreciated long-term earnings power. What looks expensive on near-term numbers can be cheap on earnings several years out.

His examples are meant to be stark. Whale Rock bought Nvidia in 2023 at what Sacerdote describes as four times earnings. It bought Tesla in 2019 for the electric-vehicle S-curve at five times earnings. It owned Apple at four times earnings. It bought Amazon when, in his view, AWS was effectively free inside the company. Earnings can move from $1 to $10, or from 50 cents to $20, “way more than you think,” if the adoption curve and competitive position are right.

The reason this is hard, he says, is that “the world doesn’t think exponentially.” Markets focus on the next quarter or year. Few investors believe they can accurately forecast two, three, or four years out. The S-curve provides a map, provided the investor also understands the moat and can model the business.

He distinguishes three questions inside the S-curve. What barriers prevented adoption during the long flat period? What changed to remove them? How tall is the curve?

The iPhone is his consumer example. Smartphones existed for years before Apple’s breakthrough. The barriers were clunky hardware, poor data networks, high prices, and limited ease of use. Apple combined a touchscreen interface, a $200 price point, AT&T’s 3G network, and an ecosystem simple enough for nontechnical users. Once those barriers fell, demand “rocketed.”

Tesla is his electric-vehicle example. The barriers were price, range anxiety, and manufacturing scale. The inflection came when Tesla reached a roughly $40,000 car, 300 miles of range, and a supply chain capable of producing at scale.

AWS is his enterprise example. At first, AWS was a hidden line item inside Amazon, covered by retail internet analysts rather than hardware or enterprise technology analysts. Sacerdote says Whale Rock realized AWS was addressing the largest total addressable market in enterprise IT because it collapsed multiple historical categories — routers, memory, storage, Dell, EMC — into one cloud infrastructure market. Initially, Whale Rock thought AWS was addressing about $600 billion in IT systems and would be 50% deflationary, making penetration only 1% or 2%. Later, Sacerdote says, they realized cloud was not as deflationary as expected: building it yourself was about the same price. That meant the TAM was larger.

Height matters because it determines holding period and exit discipline. S-curves can be dynamic, and investors must keep updating them. Electric vehicles, for example, did not play out as Whale Rock expected. The firm initially thought perhaps 40% to 50% of cars would go electric, but the curve hit a wall at 10% to 15% for various reasons. Usually, Sacerdote says, S-curves go much further, but this one required adjustment.

As a rule of thumb, Whale Rock begins to worry when penetration reaches 30% or 40%. At that point, exponential growth slows, sell-side estimates catch up, and the large earnings beats become less likely. Apple was an exception that Sacerdote says Whale Rock handled imperfectly. The firm sold in 2012 when smartphone penetration reached roughly 50% in the U.S. Apple then compounded well after a period of underperformance because its multiple fell, it added ancillary businesses, and it captured economics from applications. But Sacerdote still says the biggest Apple years were in the zero-to-50% part of the curve.

Alex Sacerdote does not present S-curve investing as a purely quantitative exercise. He cites Andy Grove’s idea that at strategic inflection points, “you can’t trust the data.” At those moments, investors need intuition, anecdotal evidence, visual pattern recognition, and what he calls whole-brain investing.

His examples are deliberately tactile. Mobile games looked limited for a long time because phone screens were small and processing power was weak. Then Sacerdote saw a 12-year-old boy in China playing a sophisticated game on a huge phone. That visual cue suggested the gaming experience had moved to mobile in a way the historical data might not yet show.

Enterprise clues are harder because the product is less visible. Sacerdote looks for demand signals at conferences and inside buyer behavior. At Gartner’s IT symposium, where tens of thousands of CIOs attend, Whale Rock saw standing-room-only sessions for Splunk and VMware during their inflection periods. With AWS, he says, the grand ballroom was packed at 9 a.m., again at 10, and again at 11. That was evidence of demand “exploding before it happened.”

Investors do not need to catch the very beginning. If the top of the S-curve is large enough, it is acceptable to miss the first year or two, or even the first 100% in a stock. Sacerdote recalls Peter Lynch telling younger investors at Fidelity to “white out the chart” because what matters is the future.

The pace of adoption depends on the technology. Whale Rock commissioned Horace Dediu, who had worked with Clayton Christensen, to study historical S-curves. Sacerdote says the firm has a wall of major adoption curves from the past century. Radio was among the fastest, reaching near-universal penetration in about seven years. Dishwashers were much slower because they had to be physically installed into homes.

That radio-versus-dishwasher contrast shapes his view of AI. B2B technologies often behave more like dishwashers: they must be integrated into existing systems, and large companies move slowly because of security, culture, and internal politics. Cloud had that problem. Companies feared putting data in the cloud until examples such as the CIA and Capital One helped change perceptions. Sacerdote says a Capital One CIO told Whale Rock the cloud was more secure, helping shift the enterprise mindset.

AI has some of those enterprise frictions, but it also has browser-based simplicity. Consumers and many workers can open a browser and use it immediately. That is why Sacerdote thinks the curve may be steeper than prior B2B technologies. The deeper enterprise integrations still matter, but the initial usage barrier is far lower.

A major S-curve is not enough. Alex Sacerdote repeatedly distinguishes between owning the curve and owning the winner inside it. The wrong company in the right curve can still go to zero.

His mobile-phone losers are RIM, Palm, Nokia, HTC, LG, and Motorola. The smartphone adoption curve was enormous, but the companies that lost competitive position did not benefit from it. He sees a similar sorting process in foundational models: around 50 companies tried to compete, and only two or three have emerged at the top.

Whale Rock’s process is to identify the S-curve, then study every company with exposure and look for one with a powerful competitive advantage. Sacerdote argues that digital businesses can have moats as strong as, or stronger than, offline businesses.

The moat categories include network effects, such as LinkedIn, Facebook, and Alibaba; industry-standard positions, such as Oracle and Bloomberg; scale, which can arrive much faster in digital markets than in physical retail; platform status, where others build on top; critical intellectual property, such as Qualcomm in mobile phones or ASML in lithography; and brand, which matters because customer acquisition cost versus lifetime value is the business model. Google, Amazon, Apple, and Elon Musk’s companies are cited as businesses that grew with powerful brands and limited traditional advertising.

Sacerdote’s AWS call in 2013 is his preferred example. At the Robin Hood Investors Conference, Whale Rock argued that “the bulls have no idea what they’re sitting on” and that Amazon had “won the war before it even started.” At the time, Whale Rock thought there was “Coke and no Pepsi.” Microsoft Azure later became Pepsi, but AWS’s lead was large enough to produce a long runway. Amazon had a seven-year lead, became an ecosystem and platform, and reached roughly 10 times the scale of competitors, making it difficult for others to match R&D investment.

For Anthropic and OpenAI, the moat question is more complex. AI is, in Sacerdote’s words, the most complex and fastest-changing S-curve Whale Rock has studied. The potential reward is high because he sees the market in the trillions — he gives a current estimate of $3 trillion to $5 trillion — but the risk is also higher.

Anthropic’s possible moats include critical IP, sustained share in code, an enterprise brand where CIOs immediately mention Claude, escape velocity in scale and fundraising, and recursive improvement from using its coding strength to improve its own model. OpenAI has a consumer franchise with “a gazillion eyeballs,” and Sacerdote says it is improving in enterprise and coding as well.

The broader internet pattern, he argues, is that the leader “goes bigger, faster, and wins.” Shopify, Amazon, and many SaaS companies are examples where the lead compounded. Exceptions exist, usually around paradigm shifts: AOL failed to transition from dial-up to broadband; Netscape came early but lacked a strong enough business model. Sacerdote does not deny that AI leadership can shift. But he says startups in Silicon Valley are building on top of the leading three model companies, and the economy is large enough for differentiated leaders to coexist.

Alex Sacerdote’s AI enthusiasm does not extend to most application software incumbents. Whale Rock once had 40% to 50% of its portfolio in software. After its April 2023 AI work, the firm initially thought large software companies might benefit: they had huge sales forces, customer relationships, proprietary data, and the ability to embed AI APIs into products. But Sacerdote says their AI products were not very good, did not move the needle, and could not command meaningful pricing. Whale Rock sold almost all of its application software exposure and entered the year net short, which he says helped in the first quarter.

His critique has several layers. AI changes the software interface. The old way of software is, in his analogy, pen and paper or horse and buggy; the new way is a jet engine, or even the transporter from Star Trek. Even if disruption is not immediate, the contrast is large enough to threaten existing models.

CIO priorities have also changed. Even if AI does not replace incumbent software, software vendors have moved down the priority list. Enterprises may prefer to spend on Anthropic tokens because the ROI is faster. That spending pressures budgets elsewhere. Software companies that previously raised prices every year may become more hesitant. Seat-based models could suffer if companies freeze hiring or reduce jobs. Sacerdote says there are smart people on both sides of the employment debate, but Whale Rock is seeing some companies “gut their jobs” or freeze hiring.

AI-native competitors may emerge. Bulls on incumbent software argue that enterprises will not build their own ERP systems, and Sacerdote says that is probably right. Old technology is sticky; mobile games did not kill console games, tablets did not kill PCs, and smartphones did not kill PCs. Enterprises prefer buying to building, and integrations matter. But it is plausible, in his view, that over one to five years, AI-native companies attack each incumbent category, the old data advantage is weakened, and AI makes replacement easier.

The incumbents’ problem is also mathematical. A large company such as Salesforce may have $40 billion in sales and perhaps $500 million or $700 million of AI ARR, in Sacerdote’s example. Even if the AI product begins working, it takes time to move the overall growth rate.

He contrasts software with AI hardware using a modified “rule of 40.” In traditional software, the rule of 40 is revenue growth plus operating margin. In AI infrastructure, Whale Rock looks at what percentage of a company’s sales are AI and what market share it has in that category. A company with 30% AI exposure and 30% category share would score 60, making it attractive under this heuristic. Most software companies, he says, have only 1% or 2% AI exposure today.

Still, he leaves room for a more positive software outcome. AI agents may make some existing platforms more important. Claude plugged into Slack could turn Slack into a key organizational repository. Systems of record such as HR, CRM, or Workday-like platforms may become the tools agents operate inside. CRM may go “headless,” meaning the human interface matters less and the AI interface goes directly into the data. That can be bearish if the incumbent becomes only a database, but it can also solidify the system if agents keep doing work inside it.

Sacerdote calls this thinking “half baked.” The direction is not settled. But his current portfolio stance reflects skepticism that traditional software incumbents are the first or cleanest beneficiaries of AI.

Alex Sacerdote’s most detailed bullish case is for AI infrastructure. His starting point is that the data center had been structurally boring for decades. Intel x86 became the dominant data-center chip architecture in the 1990s. During the cloud era, compute workloads grew 25% to 40% annually, but Moore’s Law improved at roughly the same rate. That meant the industry did not require dramatic hardware innovation. Servers, memory, printed circuit boards, enclosures, and networking components commoditized. A transition from 1 gig to 10 gig or 100 gig to 400 gig might create a temporary cycle, but the product would then commoditize again.

AI changes that because workloads are growing “10x every year” and pushing every part of the hardware stack toward physical limits. Sacerdote calls this the “de-commoditization of the hardware industry.” Unit growth is rising, but so is the need for innovation.

High-bandwidth memory is one example. Memory was once treated as a pure commodity. AI requires stacked chips, far greater input-output capacity, tight coordination with Nvidia, and multi-generation roadmaps. Sacerdote says it took Samsung years to do high-bandwidth memory well, and the component has become critical.

Celestica is another example. Historically, contract manufacturing was a poor industry: offshore, commoditized, and unattractive since the late 1990s. But Celestica retained talent from its IBM supercomputing heritage. Whale Rock noticed three years ago that it was the sole supplier of the Google TPU server and was also selling Ethernet white-box switches into cloud providers. The stock, Sacerdote says, was trading at eight times earnings.

What looked like a growth driver turned out, in his view, to be a moat. AI servers are not disposable $5,000 machines. They can be $200,000 or $300,000 liquid-cooled systems running much hotter. If one fails, the system matters. That makes the supplier, in Sacerdote’s analogy, more like a critical part provider for an aircraft: difficult to swap out once qualified. Sacerdote says Celestica proved good at liquid cooling while others failed, and retained its position.

Networking also changed. Previously, Ethernet upgrade cycles took about seven years. Now they are happening annually, which is difficult. Sacerdote says Celestica’s people helped write the open-source SONiC networking software and work closely with Broadcom. He estimates Celestica has 50% to 60% share of the cloud Ethernet switch market, a crucial category because AI is network-intensive.

Printed circuit boards show the same pattern. A regular server may require 10 layers; AI servers may require 40. Few PCB suppliers can produce them. Whale Rock also owns Elite Material, which makes copper-clad laminate, a key input into these boards. Sacerdote says PCB units are growing, layer counts are rising, average selling prices are rising, gross profits are rising, and customer visibility has changed from week-to-week ordering to four-year roadmap planning.

That changes the financial profile. A company that was once a 5% grower with low margins can become a 35%, 40%, or 50% top-line compounder for the next four years with rising margins.

Corning is another case. Sacerdote says it has an extremely high share in fiber, and he describes reading about a Microsoft data center with enough fiber to circle the world four and a half times. Corning’s fiber is thinner, more bendable, can be manufactured to precise specifications, carries higher margins, and is the fastest-growing part of its business.

He distinguishes three networking layers. “Scale out” connects server racks. “Scale across” connects data centers, especially when power constraints prevent a single giant cluster in one location. “Scale up” connects GPUs inside a rack, currently over copper and eventually, in Sacerdote’s view, over fiber. If scale-up moves to fiber, he says, Corning’s opportunity could increase by two to three times.

Power supplies follow the same logic. Every Nvidia chip or rack uses 50% to 125% more power, by Sacerdote’s estimate, which directly lifts average selling prices for companies such as Delta and Advanced Energy. He says he finds these stories almost hard to believe: ASPs rising 40% for four consecutive years, with higher margins.

The broader supply-chain claim is severe. If the AI L-curve is right, Sacerdote says, markets such as DRAM, NAND, and PCBs are already about 30% short. Even commodities can have strong cycles under those conditions. But his stronger claim is that many categories are no longer simple commodities because AI has raised the engineering difficulty and strategic importance of the components.

Alex Sacerdote calls Whale Rock the “Whale Rock Learning Machine”: a 10-person team built around accumulated technology knowledge, constant company contact, and long-running pattern recognition. The firm does 2,500 to 3,000 face-to-face management meetings a year. About 10% to 15% are with private companies.

That matters because Sacerdote believes private markets have become unavoidable in technology investing. He says the unicorn market is larger than most European stock markets, “maybe even combined,” and definitely larger than Germany or the U.K. Even before Whale Rock invested privately, the firm had to know private companies because they often were the most important competitors or category leaders.

Anthropic illustrates the process. Whale Rock saw Anthropic’s $60 billion round but passed because it did not yet know the company well, gross margins were negative, and coding had not exploded. Later, Sacerdote spent time with Dario Amodei, followed him publicly, and became impressed by the management team’s quality, focus, low turnover, code quality, and business plan. Whale Rock then built a 90-page deck using Claude Code to scour the internet for feedback on the coding market, Anthropic’s strengths, and areas for improvement. The company welcomed Whale Rock into the round, and Sacerdote says the firm “punched above our weight” in allocation.

Stripe was the earlier private-market template. Whale Rock owned Adyen, a next-generation payments company, and concluded that it could not understand Adyen without understanding Stripe “like the back of your hand.” In diligence calls with roughly 200 Adyen customers, Whale Rock asked about Stripe as well. Sacerdote came to view the two as Coke and Pepsi. During COVID, a friend with access to Stripe stock called about a sale. Whale Rock did not have full financials, but it knew enough: disclosed payment volume over half a trillion dollars, Adyen take rates around 25 to 30 basis points, Stripe likely at 40 to 50, and enough employee information to estimate profitability. Sacerdote says Stripe’s actual numbers were better than Whale Rock’s underwriting assumptions, and the firm upsized into a $100 million block.

The same research system applies in public markets. Sacerdote ties it to Philip Fisher’s “scuttlebutt” approach from Common Stocks and Uncommon Profits: talk to suppliers, customers, competitors, and management teams; identify the characteristics of leading companies; and build conviction before the market fully understands the story.

AI helps, but he is clear about its limits. It can accelerate learning in complicated areas such as ABF substrates or PCBs. It can help write notes, review quarters, and produce better reporting. But it cannot pick stocks. The analyst still needs the top paragraph: the judgment about what changed, what it means for the thesis, and whether the future is different. “Don’t just be a reporter,” he says. AI can be a great reporter, but it cannot yet do the forward-looking insight work.

His example is AppLovin. Two Whale Rock analysts, Michael and Sam, understood the ad-tech market, followed the company when it was private, knew competitors and terminology, attended industry conferences in Las Vegas and Cannes, talked to scores of people, built the model, and developed a relationship with CEO Adam Foroughi. Sacerdote says he does not see AI doing that.

External investors also play a role. Sacerdote describes a “tripod” of conviction: he likes an idea, his analyst likes it, and an outside investor he respects also likes it. Philip Fisher, he says, recommended knowing 10 or 15 like-minded investors around the country and sharing ideas. Sacerdote values those relationships as friendships and as part of the research process, provided the exchange is two-way.

Whale Rock’s product lineup grew around that research base. The firm began as a long-short fund, launched a long-only product in 2020, formalized an option for private investments before making its first private investment in 2020, later offered a hybrid fund with more private exposure, and recently launched a mega-cap tech fund. Sacerdote’s argument for the newest product is that many large pools of capital are structurally underweight the world’s largest technology companies, despite those companies often having the strongest moats and, in some cases, the most direct exposure to major S-curves.

Alex Sacerdote’s bullishness is not presented as risk-free. The first concern he names is political and social resistance. He says there is broad negativity about AI in the general population and in parts of government. As examples of that risk, Sacerdote says Maine “just banned data centers” and says only 20% of people are optimistic about AI. Negative regulation could slow the buildout. Still, he thinks “the genie is out of the bottle.”

The second risk is that model improvement slows. Sacerdote says there is already significant AI adoption available even if models stopped improving. But if leading models stop advancing, open-source models could catch up, and the model layer could become a race to the bottom. That would likely be bad for model-company equities, though he says chip companies may still benefit because they do not care who burns the tokens. Jensen Huang, he notes, wants open source to take off and emphasized it at GTC.

The third risk is that one or two major players falter and no longer need the compute they had planned to use. If AI demand is large enough, another buyer may absorb the supply. Sacerdote gives, as an example, Oracle canceling a big deal and Meta stepping in. But if a player such as Meta decided AI was not worth the resources, that would release a significant amount of expected compute demand. Whale Rock watches that carefully.

The fourth risk is simply the speed of change. AI is the fastest-changing curve Sacerdote has seen, and the model layer is especially complex. Competitive advantages that look strong need to be re-evaluated continuously.

Those risks also explain why Whale Rock has not aggressively pursued the application layer. Applications usually come later: the iPhone’s app ecosystem took several years to mature. In AI, the boundary between foundational model and application remains unclear. Applications may struggle to build enough of a moat if model providers expand into their functions. There will be great application companies, Sacerdote says, but the ecosystem is not yet clear enough for Whale Rock in many cases.

He cites Bret Taylor’s Sierra as an example of a company he is watching closely but is not involved in. Taylor’s background — Salesforce CEO, Google Maps creator, Facebook CIO — makes Sierra a serious test case for where “the rubber hits the road.” The question is whether such companies can become large, sustainable businesses.