The posts on this blog test a hypothesis. This page draws the operational inference: what the evidence implies if you are running a business that exists, or building one that doesn't yet. The analysis is only as current as the hypothesis tracker — both update together. ← Hypothesis tracker
Defending What You Have
Know which quadrant you're in
AlixPartners scored 500 PE-backed software companies on AI disruption risk. 25% are highly vulnerable; only 14% have strong moats — defined as genuine data depth combined with vertical specialisation tight enough that switching is structurally painful. The distribution isn't random: point-solution tools with generic data models and horizontal scope cluster in the exposed quadrant; systems of record with compliance obligations and proprietary data context cluster in the fortress.
The useful exercise isn't asking whether AI will affect you — it will — but locating yourself on that grid honestly. Vulnerable companies in the AlixPartners model face a $40B debt wall maturing in 2028 against revenue lines projected to shrink 15–35%. That's not a technology problem; it's a capital structure problem that technology is accelerating. Knowing your quadrant early is the difference between a managed repositioning and a forced one. Two months of vendor disclosure have added a second axis to the same grid. Intuit's Q3 FY26 print on May 20 paired a 17% workforce cut (3,000 of 18,200 employees) with multi-year Anthropic and OpenAI deals that route TurboTax, QuickBooks, Credit Karma, and Mailchimp behind Claude and ChatGPT — flagship vertical SaaS with deep data (90 million tax returns, IRS-grade compliance) voluntarily moving its UI behind an assistant front door it does not control, on the explicit bet that the data and expert layer below the UI holds enough value to compensate. The stock entered the print down 41% year-to-date; the market is not yet convinced. The refinement matters: data depth and vertical specialisation are necessary for the fortress but not sufficient. The customer segment's switching-cost endowment is the second axis. Palantir's enterprise ontology customers carry five-year integration debts that make the front door sticky; a small-business owner asking Claude for an invoice template carries none. The same fortress thesis tested at the consumer/SMB layer produces a thinner-walled version, where the data moat may hold margin but the seat definition collapses faster. When locating yourself on the AlixPartners grid, plot the second axis: how much of your customer base would choose an assistant-mediated alternative tomorrow if one existed.
Evidence: The Great SortingRe-price before the market forces you
AI inference costs are structural, not transitional. When your product relies on AI to deliver outcomes, gross margins compress from 80–90% toward 50–60% — not because you're inefficient, but because COGS now scales with usage in a way seat licences never did. Intercom moved to $0.99 per resolved conversation and grew from $1M to $100M ARR. Freshworks charges $0.10 per session. The difference in those numbers encodes a bet about whose model survives.
The seat model breaks arithmetically once AI does the work: fewer humans doing the work means fewer licences, regardless of whether your product is any good. The companies that will survive the repricing cycle are those that move to consumption or outcome pricing before the compression shows up in renewal conversations. Deloitte forecasts 40% of enterprise SaaS contracts will include outcome-based elements by 2030. The question is whether you arrive at that number on your terms or theirs. GitHub Copilot's June 1 2026 transition is the cleanest pure-pricing-model data point of the cycle so far. Microsoft kept every base subscription tier at the same nominal price ($10 Pro, $39 Pro+, $19 Business per user, $39 Enterprise per user) and converted what those fees buy from flat-unlimited usage into a fixed monthly allotment of AI Credits — $0.01 per credit — with token-metered overages above the ceiling and no fallback to a cheaper model once credits exhaust. The CPO framed it as 'aligning pricing with actual compute consumption' because 'Copilot is not the same product it was a year ago — it now powers far more complex, agentic workflows that consume far more compute.' Translated, that is a single-sentence admission that the flat seat could not absorb agentic compute at the heaviest tier of usage, and Microsoft chose to surface the variable cost as overage rather than absorb it. The arrival path is the opposite of Workday or Salesforce — those started seat-based and added consumption credits on top; GitHub Copilot started flat-unlimited and added credit ceilings underneath — but the destination is identical: flat base plus metered overage. The operational lesson for any incumbent still on flat-fee pricing is that the model only survives if heavy users are either rare enough to subsidise out of the base fee or willing to pay the overage when surfaced. Microsoft did its repricing on its own terms, on its own date, with a six-week notice window and a three-month promotional credit buffer for existing Business and Enterprise customers — exactly the 'self-funded, advance-of-the-market' posture this entry has documented at Atlassian and Workday from the other direction. Run the repricing while you still get to choose the timing and the buffer.
Evidence: The Cost Problem That Broke the Pricing ModelOwn the workflow or own the data — pick one
Snowflake grew revenue 30% full-year. Palantir grew 70% in Q4. Workday fell 40% on headcount-linked licence compression. The divergence is structural: companies that own the system of record, the compliance layer, or the proprietary data context that agents need to function are holding value. Point solutions sitting underneath an orchestration layer — where any sufficiently capable agent can route around them — are not.
The trap is trying to be both. A company that positions itself as the orchestration node and as the best-in-class point solution ends up defending two fronts with the resources of one. The mechanism is now explicit on earnings calls: ServiceNow's CFO told Q1 2026 investors that customer AI budgets are sourced from 'labor budgets coming down,' 'reallocating technology spend,' and 'eliminating more point solutions and really leaning into platform consolidation.' Fifty percent of ServiceNow's net new business now runs through non-seat pricing, and Now Assist customers spending more than $1M annually grew 130% year over year. The platform consolidators are stating directly that their AI revenue is partly funded by killing point solutions inside the customer stack. The M&A market is now pricing the same principle directly: SAP committed €1 billion over four years to acquire Prior Labs in May 2026 — an 18-month-old German AI lab specialising in tabular foundation models — with more than half a billion paid in cash up front. The deepest-data vendor in enterprise software is paying premium rates to acquire the AI capability optimised for the data shape it already owns, on the explicit thesis that enterprise AI value will accrue at the structured-data layer rather than the conversational interface. Pick the layer where your data or workflow gravity is genuinely irreplaceable, and concentrate there before the consolidator earnings cycle decides for you — or before the M&A market reprices the asset out of reach. The May 27 2026 Q1 FY27 prints from Salesforce and Snowflake collapsed the optionality further. Salesforce reported Agentforce ARR of $1.2 billion, up 205% year over year and roughly 50% sequentially from the $800 million figure at the end of January, with combined Agentforce and Data 360 ARR near $3.4 billion (+200%+); Marc Benioff used the call to name the next strategy 'Headless 360' — expanding the addressable market into surfaces never previously monetized, meaning the company will sell its data and orchestration layer to be called from wherever the customer happens to be working rather than only on the Salesforce UI. Snowflake the same day posted product revenue of $1.33 billion (+34%), the strongest sequential dollar growth in company history, with 13,600+ accounts now using AI capabilities, Snowflake Intelligence accounts doubling quarter-over-quarter, and Sridhar Ramaswamy positioning the company explicitly as the 'control plane for the Agentic Enterprise.' Workday's print six days earlier ran the same playbook from a different starting point: ~$500 million annualised agentic AI revenue, customer agent count doubled quarter-over-quarter, operating margin expansion 1,154 basis points year-over-year. Three different vendor architectures executing the same trade in the same week: own the data, own the control plane, sell the orchestration layer to whichever agent surface the customer prefers. The data-and-workflow combination is now the explicit strategy at the largest pure-play CRM, the largest pure-play data cloud, and the largest pure-play HR vendor — not a niche fortress position but the converging playbook of the canonical horizontal SaaS incumbents. For builders deciding where to concentrate, the implication is that the 'own the workflow OR own the data' framing is becoming 'own the workflow AND the data, and price the agent traffic that traverses them'; whichever side you concentrate on, plan for the other to be expected by buyers within twelve months.
Evidence: The Dispatch LayerFix your distribution before AI kills it
Monday.com's $500K+ ARR enterprise cohort grew 74% year-on-year. Its SMB self-serve motion collapsed. The cause wasn't AI replacing the product — it was Google's AI-powered search results replacing the organic traffic that drove self-serve signups. Co-CEO Roy Mann told investors that the cost to acquire and expand self-serve customers had risen sharply, with returns below historical levels. The product still works. The customers stopped arriving.
This is the least-discussed existential risk in the current cycle: AI disrupting distribution economics before it disrupts product value. If your acquisition model depends on SEO-driven search traffic, content-driven signups, or any channel where an AI summary can answer the intent that used to drive a click, the threat isn't a competitor building a better product — it's the channel itself evaporating. The audit is simple: map where your new customers come from, and ask whether those sources still exist in a world where AI answers the query.
Evidence: The Phantom RepricingSelf-fund the pivot before the market forces it
Atlassian cut 1,600 employees — 10% of its workforce — on March 11, 2026, with the announcement explicitly framing the layoff as a way to 'self-fund further investment in AI and enterprise sales.' Six weeks later, Q3 FY26 revenue grew 32% year over year to $1.79 billion against expectations of $1.69 billion, cloud accelerated to 29% growth, full-year guidance was raised, and the stock rallied roughly 30% on the day. The $225–236 million restructuring charge landed mostly in the same quarter the market rewarded. Rovo, the agent SKU stapled on top of the seat product, grew credit usage more than 20% month over month, with adopters expanding ARR at roughly twice the rate of non-adopters. Going into the print, the stock was already down about 45% year to date.
The order of operations is the lesson, not the layoff itself. A seat-based incumbent with genuine workflow gravity has a finite window to fund the transition from a position of strength: cut human op-ex while revenue is still growing fast enough to absorb the restructuring charge in a single quarter, and direct the proceeds at a credit or outcome SKU layered on top of the existing seat product rather than adjacent to it. Price the bundle so it pulls usage upward — Atlassian's Teamwork Collection ships ten times more Rovo credits than the standalone tier, and bundle customers use three times more credits per user. The 45% year-to-date drawdown going into the print is the warning condition: the market will discount the un-pivoted vendor faster than the un-pivoted vendor would discount itself, and acting late means restructuring under duress, with the proceeds servicing debt rather than funding product. The Workday Q1 FY27 print on May 21–22 2026 shows the same playbook executed without the layoff. Returning co-founder CEO Aneel Bhusri signed the 10-Q on his first day back, set an explicit FY27 target of keeping headcount as close to flat as possible relative to Q1 levels, and raised the non-GAAP operating margin guide by 50 basis points to 30.5% on the same release. GAAP operating margin expanded from 1.8% to 13.3% year over year — a 1,154 basis point swing in twelve months — while agentic AI annualised revenue reached roughly $500 million with the customer agent count doubling quarter-over-quarter and new ACV from agentic AI growing more than 200%. The mechanism is the layoff-less variant of the Atlassian template: instead of cutting headcount outright and pricing the restructuring into a single quarter, hold headcount flat and let AI absorb the productivity gain that would otherwise have been delivered by hiring. The vendor does to itself, internally, what its customers are doing externally — and the 10-Q is explicit that the customers are doing it: management characterises recent quarters as showing moderation of revenue growth from deal scrutiny, lengthened sales cycles in net new opportunities, and reduced growth in headcount-level commitments at the renewals of existing customers, with federal-funding-tied verticals (government, higher education, healthcare) explicitly identified as the locus of pressure. The market priced the substitution monetisation ahead of the renewal admission, +10% on the day — same direction as Atlassian's reaction, opposite sign to ServiceNow's beat-the-quarter reaction four weeks earlier. The pattern is clarifying: vendors whose agent line is large enough and growing fast enough to look like genuine substitution capture get rewarded even when the seat-renewal admission is explicit; vendors whose agent line is fast but small relative to seat exposure do not. The order of operations stands; the cost-structure path branches into layoff-funded (Atlassian, Cloudflare, Cisco) and flat-headcount-funded (Workday) variants, and Workday's print is the cleanest evidence to date that the second variant works in the same direction as the first.
Evidence: The Self-Funded PivotWhat the Hypothesis Opens Up
Target the labor budget, not the software budget
Sequoia's framing is precise: every dollar spent on software corresponds to six dollars spent on services. AI autopilots don't compete with the software budget — they compete with the services budget. The target markets are outsourced verticals where human labour delivers outcomes at predictable per-unit costs: insurance claims processing, healthcare revenue cycle management, legal document review, accounting workflows. These markets are large ($50–200B in identifiable outsourced spend), they price on outcomes already, and they're not defended by the same switching-cost moats that protect SaaS systems of record.
The business model writes itself from the economics: charge per resolved claim, per processed document, per completed reconciliation — and deliver at a cost structure that legacy BPO providers cannot match. The constraint isn't AI capability; it's regulatory tolerance and audit trail requirements, which are solvable engineering problems. The companies building here aren't SaaS companies; they're service businesses with software economics.
Evidence: Selling the Work, Not the ToolBuild in the fortress quadrant from day one
The AlixPartners data shows only 14% of existing software companies sit in strong-moat territory — deep data combined with vertical specialisation. Most niche verticals with genuine data complexity haven't been claimed by an AI-native builder. The advantage of starting now rather than repositioning is that you can architect for data depth from the first line of code, rather than retrofitting it onto a horizontal tool that was never designed to hold proprietary context.
Palantir's Q1 2026 print, covered in 'Above the Model Layer,' gives the financial signature this quadrant produces when AI ships into it correctly: revenue growing 85% year over year (US commercial +133%), adjusted gross margin at 88%, adjusted free cash flow margin at 57%, net dollar retention 150%, customer count expanding 31% to 1,007, 47 deals at $10M+ in a single quarter, and a Rule of 40 score of 145% — matched in public markets only by NVIDIA, Micron, and SK Hynix. Every variable the broader hypothesis tracks moves the opposite direction at this layer, and the mechanism is specific: when the model layer commoditises (token costs down ~1000x since 2023, by Palantir CTO Shyam Sankar's account), the value released by that commoditisation flows up the stack to whichever layer holds the customer-specific context AI agents have to query before they can act. The playbook is built around that cleavage point: pick a vertical where the data model is genuinely complex and hard to replicate externally, embed deeply enough in the workflow that the customer's processes are encoded into the platform itself, and price on the depth of that integration rather than seats or tokens. The companies in this quadrant trade at premiums the exposed segment will never recover; the financial signature is observable now, not theoretical.
Evidence: The Great SortingThe $40B distressed stack
PE-backed software faces a specific capital structure problem: $40B in debt matures in 2028, concentrated in companies whose revenue lines are projected to shrink 15–35% before that maturity arrives. AlixPartners flagged the mechanism explicitly. IGV closed at 3.6x forward EV/Sales — the lowest since 2011 — and the public valuation compression is already flowing into private credit markets. Some of this software is genuinely broken. Some of it is exposed-but-fixable: good underlying workflow value, wrong pricing model, wrong cost structure.
The opportunity in the distressed segment is acquiring businesses with real workflow gravity — customers who renew because switching is painful, not because the product is loved — and restructuring them around AI-native delivery. The target profile: high net revenue retention, low NPS, high COGS, seat-based pricing, horizontal market. The intervention: rebuild delivery on AI, move pricing to outcomes, cut the cost structure accordingly. The debt wall creates a time-bounded acquisition window between now and 2028.
Evidence: The Great SortingThe governance and trust layer
Microsoft open-sourced the Agent Governance Toolkit covering all OWASP Top 10 agentic risks — auditability, access control, prompt injection, data leakage, liability attribution. That Microsoft shipped this as open-source is a signal: the capability problem is largely solved; the trust and compliance infrastructure around it is not. Enterprise adoption of AI agents is currently blocked less by what agents can't do than by what enterprises can't audit, certify, or defend in a regulatory context.
The market being created here isn't governance tooling for its own sake — it's the infrastructure that unlocks enterprise AI agent adoption at scale. Every agent deployment needs an audit trail, a compliance boundary, and a liability model. None of that ships in the model. The companies that build the monitoring, certification, and governance layer around agent deployments will occupy a position structurally similar to what security vendors occupied in the early cloud transition: not optional, not fungible, and priced accordingly.
Evidence: The Week the Market Said It Out LoudRent the capex, don't carry it
Microsoft and Meta cut roughly 17,000 staff between them on the same Thursday in April 2026 — Microsoft's first voluntary buyout in fifty-one years — to fund $115–135B and $145B in 2026 AI capex respectively. Combined Big Four AI infrastructure spend approaches $700B in 2026. Whether or not those bets earn their depreciation curves, the capacity is being built, and the depreciation schedules force aggressive utilisation pricing. Inference cost per unit is going to be under structural downward pressure as long as the hyperscalers compete for fill rate.
The implication for AI-native builders is to design for cheap, abundant inference and resist the temptation to own infrastructure. Build on top of capacity that Big Tech is overbuilding, ride the subsidised COGS curve while it lasts, and concentrate capital on application-layer differentiation — data context, workflow embedding, distribution. The supply side of that capex now shows the same margin signature. Cisco's Q3 FY26 print on May 13 raised its FY26 hyperscaler AI infrastructure order guidance from $5B to $9B — 4.5x FY25 — and posted non-GAAP product gross margin down 330 basis points to 64.3% on the same mix shift, while the company simultaneously announced 4,000 layoffs and up to $1B in restructuring charges. The picks-and-shovels position is not a margin-rescuing destination for vendors fleeing softening software economics: it captures volume on a per-unit margin that hyperscaler buyers actively compress. The window is finite: this capex cycle ends either in consolidation among hyperscalers or in a capacity rationalisation that re-prices inference upward. The right strategy now is the wrong strategy in 2030. Plan for both — and price equipment-vendor equity for the margin tax, not just the revenue line.
Evidence: The Capex CutStay portable when the providers can't
Microsoft's exclusive cloud arrangement with OpenAI was the most valuable AI distribution moat in software. It lasted six years and nine months. The morning after the renegotiation, AWS shipped GPT-5.5, GPT-5.4, OpenAI Codex, and Bedrock Managed Agents into limited preview. Microsoft kept its 27% equity stake and OpenAI's $250B Azure spend commitment through 2032 — but the exclusive distribution rights that justified the original investment thesis were arbitraged away in a single news cycle. The model layer is fungible faster than the cloud layer; the cloud layer is fungible faster than the application layer.
The operational implication for builders is portability as a first-class design constraint. Use abstraction layers — Bedrock, Vertex, Azure OpenAI, OpenRouter — that let you swap model providers without rewriting product surface area. Negotiate cloud commitments that don't lock you to a single AI distribution stack. Treat any 'exclusive partnership' your suppliers offer as time-bounded by definition. The investor-facing version of the same lesson sharpened in May 2026, when Anthropic committed $200B over five years for 5GW of TPU capacity on Google Cloud — roughly 40% of Alphabet's $462B Q1 backlog. Combined Anthropic and OpenAI contracts now account for nearly half of the roughly $2 trillion of long-term backlog disclosed by the four largest hyperscalers (OpenAI alone: $250B Azure, $300B Oracle over five years, $38B AWS, $22.4B CoreWeave). The financing loops: Alphabet has agreed to invest $10–40B in Anthropic; Anthropic commits $200B back to Alphabet. Anthropic's $30B run rate as of April 6 implies it must scale 20–30x by 2029 to honour the contract. Oracle lost roughly half its market cap over five months once investors absorbed how much of its backlog was OpenAI-linked. The cloud-AI fortress is real only conditional on two cash-burning private labs growing into commitments larger than their current revenue. The companies that survive multiple repricing cycles are the ones whose product economics improve when foundation models become substitutable and whose investment theses do not depend on the cloud layer's headline backlog being durable. Microsoft's 20M paid Copilot seats and $37B AI revenue run rate are the proof that application-layer value capture is real; the dissolution of the exclusivity, and the circular structure of the backlog underneath it, are the proof that the layers below are not where you should plant your flag — as a builder or as an investor.
Evidence: The Six-Year MoatLast updated: 2026-06-01 · hypothesis tracker