Why AI Is Still in the First Inning

S-Curves Drive Exponential Earnings, Not Linear Growth

When a technology hits the right part of its adoption curve, unit growth and earnings grow exponentially, not linearly. Most investors focus on next quarter; those who understand S-curves can predict 2–4 years ahead and buy best-in-class companies at 4–5x earnings when the market misses the inflection.

4–5x earnings

Entry valuation for best companies at S-curve inflection

Nvidia (2023), Tesla (2019), Apple, Amazon AWS all bought at 4–5x earnings before exponential growth.

Every Technology Has a Long Flat Tail Before Takeoff

Smartphones existed 10 years before the iPhone; the internet 20 years before Netscape; electric vehicles 15 years before Tesla's 2019 inflection. Barriers to adoption (price, performance, ecosystem, network) must fall away before the tornado of demand hits.

Barriers to adoption must fall before exponential growth. iPhone removed price, interface, and network barriers.

Know the Height of the S-Curve to Know When to Sell

Understanding the total addressable market (TAM) and ultimate penetration tells you how long growth will last. AWS addressed $600B of IT systems; cloud turned out to be non-deflationary, expanding the TAM. Exponential growth typically stops around 30–40% penetration.

30–40%

Penetration threshold where exponential growth ends

Beyond this point, sell-side catches up, beats become predictable, multiples compress. Apple sold at ~50% US smartphone penetration in 2012.

B2B Adoption Is Slower Than Consumer (Dishwasher vs. Radio)

Consumer tech (radio) can reach 100% in 7 years; B2B tech (dishwasher, cloud, SAS) takes much longer because it requires integration with existing systems, security vetting, and cultural buy-in from IT and management. AI is different: it's browser-based, no integration needed, so adoption is faster.

Consumer tech adopts fast; B2B requires integration and security vetting. AI skips integration, so it's faster than cloud.

Enterprise AI Is Less Than 1% Penetrated—Straight Up

Unlike typical S-curves with a slow start, enterprise AI adoption is so steep it's called an L-curve: nearly vertical from the start. Anthropic has 14–15 million DAUs, but only ~10 basis points (0.1%) of knowledge workers are truly using AI agentic capabilities. The jump to 1–3% will trigger a light-switch moment in enterprise.

<1%

Enterprise AI market penetration

Called an L-curve: nearly vertical from day one. Compare to cloud at 10% infrastructure penetration; AI is even earlier.

Coding Is the True Unlock—$500B Market from Coding Alone

In 2025, coding tools (Claude Code, GitHub Copilot) shifted from writing 20% of code to being nearly fully agentic. Anthropic employees spending $100/day on tokens = $30k/year per user. With 20 million coders globally, that's a $600B revenue opportunity on 7–9 month old technology, before broader enterprise adoption.

$600B

Addressable market from coding alone

20M coders × $30k/year = $600B. This is before general enterprise AI adoption and before models improve further.

Anthropic's Competitive Moat: Enterprise Focus, Coding Leadership, Ecosystem

Anthropic differentiated by focusing purely on enterprise (vs. OpenAI's consumer play) and dominates coding. Unlike cloud (commodity servers), AI models have critical IP: training methods, skills, and differentiation. Anthropic is building an ecosystem (SDK, orchestration, harness) around the API, similar to AWS's early lock-in strategy.

Anthropic's moat: not just the model, but the ecosystem and enterprise focus. Similar to AWS's early dominance.

Three-Horse Race: Anthropic, OpenAI, Google (Gemini)

Of ~60 foundational model startups in 2023, almost all have died. Amazon and Meta's efforts faltered. The market is consolidating to an oligopoly of 3 leaders with superior token quality. Being 80% close to the top is not enough; the gap from 80% to 85% is a huge unlock. Open-source can't leapfrog because it lacks compute.

Winner-take-most dynamics: 60 startups → 3 leaders. Superior token quality and compute create durable moat.

Compute Shortage: Already 30% Short; Demand Growing 10x/Year

AI workloads grow 10x annually, pushing every hardware component to physical limits. Anthropic has only half the compute it needs. DRAM, NAND, PCB markets are already 30% short. This is not a future problem—it's happening now. Every layer of the stack (power, cooling, networking, memory) is constrained.

30% short

Current shortage in DRAM, NAND, PCB, power supplies

Demand growing 10x/year vs. Moore's Law improvement of ~25–40%/year. Shortage will persist for years.

Decommoditization of Hardware: Innovation Across Every Layer

For 40 years, data center hardware was commoditized (Intel x86, standard servers). AI changes everything: high-bandwidth memory stacked 10 chips high, input/outputs 10x faster, liquid cooling, 40-layer PCBs (vs. 10 for old servers), $300k AI servers (vs. $5k commodity servers). Every component now requires innovation and commands premium pricing.

Every layer of the AI server stack is being re-innovated. This is a renaissance for hardware companies.

Celestica: Contract Manufacturer Turned AI Infrastructure Winner

Celestica was a commodity contract manufacturer (disaster industry since 1999). But it retained IBM supercomputing heritage and became sole supplier of Google TPU servers. It also dominates Ethernet white-box switches (50–60% cloud market share). Liquid cooling, software (open-source SONiC), and Broadcom partnership create durable moat. Stock was 8x earnings; now 50–60% CAGR for 4+ years with rising margins.

From commodity to critical: sole supplier of Google TPU, 50–60% Ethernet switch share, liquid cooling expertise.

Elite Materials & PCB Suppliers: 50–60% Unit Growth + Rising ASPs

AI servers need 40-layer PCBs (vs. 10 for commodity servers). Few suppliers can make them. Elite Materials makes copper clad laminate (key ingredient). PCB units growing 50–60%, layer counts rising, ASPs rising, gross margins expanding from 5% to 35–50%. Visibility shifted from 'call us next week' to '4-year roadmap partnership.'

PCB suppliers: unit growth + ASP growth + margin expansion + long-term visibility. Compounding at 50–60% CAGR.

Corning Fiber: 4.5x World's Fiber in One Data Center

Corning makes specialty fiber for data centers—thinner, more bendable, custom-spec'd, higher margin. One Microsoft data center contains enough Corning fiber to circle the world 4.5 times. Fiber is fastest-growing part of Corning's business. Scale-out (connecting racks) and scale-across (connecting data centers) are driving demand; scale-up (GPU-to-GPU within rack) will eventually move to fiber, 2–3x-ing Corning's opportunity.

4.5x

World's circumference of fiber in one Microsoft data center

Corning's specialty fiber is highest-margin, fastest-growing business. Scale-up migration to fiber will 2–3x opportunity.

Power Supplies & Advanced Energy: 50–125% More Power, Rising ASPs

Each Nvidia GPU/rack uses 50–125% more power than legacy servers. Delta and Advanced Energy supply power; ASPs rising 40%/year for next 4 years with higher margins. This is not a one-time bump—it's structural and sustained.

40%

Annual ASP increase for power supplies (4+ years)

50–125% more power per GPU drives structural demand. Higher margins, multi-year visibility.

Software Companies' AI Products Aren't Moving the Needle

Large software incumbents (Salesforce, Workday, etc.) tried to bolt AI onto existing products. Results were poor: products not differentiated, no one would pay extra, no revenue lift. Whale Rock was 40–50% software 5 years ago; now net short. The problem: AI is faster ROI, so CIOs are spending on Anthropic tokens instead of software licenses.

Software incumbents' AI products failed to differentiate or generate revenue. CIOs prefer direct Anthropic tokens.

Four Headwinds for Legacy Software

1) CIOs' priority list for software has dropped; they're spending on AI tokens instead. 2) Budget pressure: AI spending crowds out software budgets. 3) Price power eroding: software companies can no longer raise prices annually. 4) Hiring freezes: fewer seats sold if companies are automating headcount.

Legacy software faces structural headwinds. Even if AI doesn't disrupt, these dynamics hurt growth.

AI-Native Companies Could Disrupt Each Vertical

In 1–5 years, AI-native startups could build ERP, CRM, HCM replacements from scratch. Data advantage of incumbents could be negated. Switching costs lower with AI-driven deployment. Risk: incumbents have huge installed bases and integration, but the threat is real and growing.

One Exception: Network-Effect & System-of-Record Tools

Slack, Workday, CRM systems could become more valuable if AI agents run on top of them. If agents operate inside Slack or CRM as a human would, these tools become permanent fixtures. Headless CRM and agent-native interfaces could preserve incumbents' lock-in.

Anthropic Investment: From $180B Valuation to $9B Potential

Whale Rock invested in Anthropic at $180B valuation in August 2025. The firm created a 90-page PowerPoint deck using Claude Code to research the coding market and Anthropic's competitive position. They convinced the company to let them in by demonstrating deep knowledge and staying close to the CFO. The investment thesis: coding market alone is $600B+; Anthropic is winning; escape velocity reached.

$180B

Anthropic valuation at Whale Rock's investment (Aug 2025)

Valuation jumped from $100B to $1B to $9B+ in months. Whale Rock's conviction: coding market is $600B+; Anthropic leads.

Stripe Investment: Deep Due Diligence on Payments Market

Whale Rock's first private investment (2020). They analyzed Stripe vs. Audon (their public holding). Stripe had $550B TPV at $35B valuation; Whale Rock calculated take rate (40–50 bps vs. Audon's 25–30 bps) and employee count to estimate profitability. Actual TPV was closer to $1T. They upsized from seller to $100M block. Stripe's willingness to have long-term holders (vs. VCs who sell) helped.

Whale Rock's due diligence: TPV was 2x disclosed; take rate 70% higher than competitor. Upsized to $100M.

How to Get Allocation in Private Rounds: Know the Company, Build Relationships

Whale Rock does 2,500–3,000 face-to-face meetings with management teams per year (~10–15% with privates). They spend time with founders, listen to podcasts, meet with CFOs. For Anthropic, they created a 90-page research deck to show conviction. For Stripe, they had deep knowledge of the payments market from owning Audon. Relationships and demonstrated expertise open doors.

Network Effects: LinkedIn, Facebook, Alibaba

Network effects create powerful moats: the more users, the more valuable. These are durable and hard to disrupt.

Industry Standard: Oracle, Bloomberg

Become the standard and lock in customers. Oracle has all the database admins, all the tuned software. Bloomberg is the financial terminal. Switching costs are prohibitive.

Scale & First-Mover Advantage: AWS

AWS got 7-year lead on competitors. Scale allowed R&D investment that competitors couldn't match. Became platform; ecosystem built on top. Amazon won before the war started.

7 years

AWS lead over competitors

First-mover + scale = 10x size of competitors. Competitors couldn't invest in R&D to catch up.

Critical IP: Qualcomm, ASML

Qualcomm: can't make a phone without paying them. ASML: can't make a chip without their lithography. These are chokehold positions.

Brand & Cost-to-Acquire: Google, Amazon, Elon

Google and Amazon never had to advertise. Elon never advertises. Brand creates organic demand and low CAC. This is a powerful moat.

AI Foundational Models: Multiple Modes Converging

Anthropic and OpenAI have scale, brand, critical IP (training methods), and ecosystem. They're building platforms that others build on. This mirrors AWS's playbook.

On the Internet, the Leader Gets Bigger, Faster, and Wins

Shopify, Amazon, SAS companies: once they lead, they keep leading. Compounding returns to scale, network effects, and brand. Exceptions are rare and usually require paradigm shifts (AOL/dialup → broadband; Netscape's weak business model).

Why 60 AI Startups Died; 3 Will Win

Foundational models require massive compute spend. Only 2–3 companies can afford it. Token quality matters: 80% to 85% is a huge unlock. Open-source can't leapfrog. Winner-take-most dynamics are accelerating.

60 → 3

Foundational model startups (2023) → survivors (2025)

Compute cost and token quality create winner-take-most. Only 3 can survive.

Risks to Anthropic & OpenAI: Model Improvement Slowdown, Regulation, Compute Shortage

If models stop improving, open-source catches up and it's a race to the bottom. Negative regulation or public sentiment could slow adoption. If compute becomes unavailable, growth stalls. But if AI is as big as expected, someone else will absorb unused compute.

AI Can't Replace Judgment; It Augments Scuttlebutt

Whale Rock uses Claude Code to write notes, review quarters, and get up to speed on new areas (PCBs, substrates). But AI can't pick stocks or develop conviction. The job is still: meet with 2,500–3,000 management teams/year, talk to competitors, customers, suppliers. AI is a great reporter; humans provide wisdom and predict the future.

The Whale Rock Learning Machine: 20 Years of Compounded Knowledge

10 highly experienced analysts (avg. 10+ years), with Andrew and Michael at 19 and 18 years respectively. They've covered multiple tech cycles (internet, mobile, cloud, e-commerce, AI). Consistency + depth = ability to see patterns others miss. Same team covers public and private.

20 years

Compounded knowledge at Whale Rock

10 analysts, avg. 10+ years experience, multiple tech cycles. Consistency beats turnover.

Product Evolution: Long/Short → Long-Only → Privates → Mega-Cap Tech

Whale Rock started with long/short (15 years). Added long-only (2020, now larger). Formalized privates (2015, broke seal 2020). Hybrid fund (2021, 80% privates). Mega-Cap Tech Fund (2025, top 30 global market caps, pick 12–13). Each product serves different LP needs while leveraging same research machine.

Product evolution driven by LP demand and research capability. Same team, different expressions.

Why Mega-Cap Tech Fund: Endowments Massively Underweight Large Tech

Endowments have huge private allocations, limited public, and often underweight large-cap tech (belief: no alpha in large-cap). But in digital economy, leaders get bigger, faster, and win. Whale Rock's Mega-Cap Tech Fund targets top 30 global market caps, picks 12–13 with best competitive advantages. Significant alpha opportunity.

Massive underweight

Endowments' exposure to large-cap tech

Belief that large-cap = no alpha is wrong. Leaders compound; Whale Rock's Mega-Cap fund captures this.

Negative Public Sentiment & Regulation Could Slow AI Adoption

Only 20% of people are optimistic about AI. Maine banned data centers. Negative regulation is a real risk. However, the genie is out of the bottle; adoption will likely continue despite headwinds.

20%

Public optimism about AI

Negative sentiment and regulation are risks, but adoption momentum is strong.

Model Improvement Slowdown: Race to the Bottom

If Anthropic or OpenAI hit a wall and stop improving, open-source catches up. This could trigger a race to the bottom on pricing and margins. However, chip companies benefit regardless of who wins.

Compute Shortage Could Limit Growth

If compute becomes unavailable, growth stalls. But if AI is as big as expected, someone else (Meta, Microsoft, others) will build capacity. Compute shortage is more likely to be a temporary constraint than a permanent ceiling.