Skip to content
Compound Startups AI Agents Data Infrastructure September 16, 2026

The Compound Startup
Sandwich

Own the data below SaaS and the agents above it. The product surface in the middle is now a build cost, not a moat. Here is the arithmetic.

By Mohinish Shaikh

Published: September 16, 2026  ·  Read time: ~17 minutes

Share Link copied

The thesis

On 15 September 2026, Shriram Sridharan, co-founder and CTO of Rox, published a short account of how his team spent roughly $20M over 2.5 years building what the SaaS era would have needed ten or more separate companies to build. The structural claim inside it is the interesting part: build the data foundation below SaaS and the agents above it, and leave the product surface in the middle alone until you have earned the right to replace it.

We think the claim is correct, and it is now testable. Public filings, index reports and price series from the last twelve months let you check every load-bearing assumption in it. This post does that check.

What we find: the middle layer has genuinely collapsed in cost, the foundation layer genuinely has not, and the incumbents' response so far has been a rename rather than a rewrite. The counter-case is real too, and it is mostly about deployment and financing rather than architecture. All figures are dated and sourced. Reliability notes are at the end.

Six numbers that frame the argument

305

SaaS apps in the average company portfolio (Zylo 2026 SaaS Management Index)

$19.8M

Average annual spend on licences nobody uses (Zylo, 2026)

312x

Fall in the price of GPT-4 level reasoning, Mar 2023 to Dec 2024 (Epoch AI)

2.1%

Monthly decay rate of B2B contact data, about 22.5% a year (MarketingSherpa benchmark)

1.9%

Agentforce ARR as a share of Salesforce FY26 revenue (our calculation from Salesforce Q4 FY26 results)

11%

Of enterprises actually running AI agents in production, against 79% who say they adopted them

1. The advice that expired

"If you do everything, you do nothing well" was not bad advice. It was an accurate description of a cost structure. When a product line needed a team, a roadmap, a go-to-market motion and three years, doing two of them at once really did mean doing both badly. Parker Conrad's compound startup argument at Rippling was the first serious attack on that constraint: many tightly integrated products on one shared platform, sold to one buyer, replacing a shelf of standalone tools.

Rippling's numbers say the model works. It reached $1B in annualised revenue in March 2026, up from $850M at the end of 2025, growing about 78% year over year, at a $16.8B valuation set in its $450M Series G in May 2025. It runs ten or more product lines each above $1M ARR, and new products typically clear that bar within five to six months of launch. That is the SaaS-era version of the compound startup: the shared asset is the employee record, and the products are still products.

What changed in the AI era is not the compound thesis. It is which layer of the stack is scarce. Three things happened at once on the buyer side, and all three are measurable.

The buyer is full

Zylo's 2026 SaaS Management Index puts the average organisation at 305 applications and $55.7M of annual SaaS spend (median $20.6M). App counts fell 0.07% year over year while spend rose 8%: portfolios have stopped growing, bills have not. Licence utilisation improved from 47% to 54%, which still leaves $19.8M a year going to seats nobody opens, and 61% of organisations reported cutting projects because of unplanned SaaS cost increases. Large enterprises above 10,000 employees spend between $123.5M and $375.5M a year.

No buyer in that state is looking for application number 306. The scarce resource on the other side of the table is not product surface. It is attention, integration budget and a procurement slot.

The pricing unit is breaking

Seat-based pricing fell from 21% to 15% of SaaS companies in twelve months, while hybrid models rose from 27% to 41%. A Cruxy survey of 300 SaaS CEOs in April 2026 found 97% planning to retire seat-based pricing within two years. Gartner's forecast is that 40% of enterprise SaaS will carry outcome-based components by 2026, against 15% in 2022. The reason is mechanical: if an agent does the work, the seat count stops being a proxy for value delivered, and net revenue retention loses its expansion engine.

The market has already repriced it

Median public SaaS revenue growth fell from 14.0% in Q4 2025 to 11.8% in Q2 2026, with consensus pointing below 10% during 2027. The median EV/Revenue multiple sits near 4.6x as of August 2026, against an 18.6x peak in 2021. Multiple compression of that size is not a sentiment problem. It is the market marking down the durability of a product surface.

Table 1. What changed between the two eras, and what it does to strategy.
Constraint SaaS era AI era (2026) Strategic consequence
Cost of a product lineYears, a team, a funding round3 to 6 months on a small teamSurface stops being a moat
Scarce buyer resourceBudgetAttention and integration slotsConsolidation beats addition
Unit of valueSeatCompleted workPricing detaches from headcount
Where the data livesInside the system of recordWarehouse and lakehouseThe record loses its read monopoly
Where the work happensInside the application UIAround it, in agent loopsThe UI stops being the product
What compoundsFeature depthContext depth and workflow ownershipBuild below and above, not across

2. The sandwich

The move is not "replace the SaaS tool". It is to take the two layers the SaaS tool sits between, and let the tool keep running in the middle until it is no longer load bearing. Below it: the data foundation that gives agents the context to act. Above it: the agents that complete the end-to-end work. Rippling replaced payroll outright. The sandwich does something stranger and cheaper, which is to consume the incumbent rather than displace it. Its API becomes a data source. Its UI becomes a screen nobody opens, because by the time a person would have opened it, the work is already finished.

The compound startup sandwich: three layers of the stack Three stacked layers. The top layer is agents that complete the work, which a coding agent can only partly compress. The middle layer is the SaaS product surface, which a small team can now rebuild in three to six months per line. The bottom layer is the data foundation, which cannot be compressed by tokens because its cost is calendar time and access. Context flows from the bottom layer directly to the top, bypassing the middle, and the middle layer's APIs are consumed as data sources by the foundation. REBUILDABLE BY CODING AGENTS? ABOVE SAAS Agents that complete the work deal management pre-meeting briefs post-call follow-up CRM hygiene last-minute QBR deck renewal watch Partly. The agent is code. The workflow it completes is not. By the time a rep would open the middle layer, the work is finished. THE SAAS MIDDLE The product surface sequencer dialer enrichment scoring reporting Historically, ten or more separate companies. Yes. 3 to 6 months per line on a small team with a large token budget. its API becomes a data source BELOW SAAS The data foundation entity model identity resolution event history permissions freshness and lineage write-back No. Priced in calendar time and access, not in tokens. context flows up, past the middle
Figure 1. The compound startup sandwich. The two layers a small team can defend are the outer ones. The middle is what a token budget now buys.

Read the diagram as a claim about where cost sits, not about where value sits. All three layers are valuable. Only one of them is cheap, and it is the one the last twenty years taught everyone to treat as the company.

3. Why the middle got cheap

Two independent cost curves collapsed at once. Neither of them is a forecast. Both are measured.

Curve one: the price of a capability

Epoch AI tracks the cheapest price at which a given level of model capability is available. The pattern is consistent and brutal: once a capability is reached, its price falls by one to two orders of magnitude within about two years. The price of GPT-4 level performance on PhD-level science questions fell from $37.50 per million tokens in March 2023 to $0.12 in December 2024, roughly 312x in 21 months. GPT-3 level performance on MMLU fell from $60.00 in November 2021 to $0.18 in February 2025. Across capability milestones, Epoch puts the rate of decline anywhere between 9x and 900x per year, with GPT-4 level science reasoning falling about 40x per year.

The collapse in the price of a fixed capability level, 2021 to 2025 A logarithmic chart of price per million tokens over time. GPT-3 level capability on MMLU fell from sixty dollars in November 2021 to eighteen cents in February 2025, about 333 times cheaper. GPT-4 level capability on GPQA Diamond fell from thirty seven dollars fifty in March 2023 to twelve cents in December 2024, about 312 times cheaper. GPT-3.5 level capability on MATH-500 fell from three dollars twenty five in June 2023 to eighteen cents in February 2025, about 18 times cheaper. $100 $10 $1 $0.10 price per 1M tokens (log scale) 2021 2022 2023 2024 2025 2026 $60.00 $0.18 $37.50 $0.12 $3.25 GPT-3 level, MMLU: 333x cheaper GPT-4 level, GPQA Diamond: 312x cheaper GPT-3.5 level, MATH-500: 18x cheaper
Figure 2. Price of a fixed capability level over time, log scale. Source: Epoch AI, LLM inference price trends. The two green and blue endpoints are 21 to 39 months apart. Note that the MMLU and MATH-500 series converge on the same $0.18 point in February 2025.

Curve two: who writes the code

The second curve is the share of shipped code that a model wrote. Measurements differ by instrumentation, which is exactly why the spread is informative: every method, at every scale, lands somewhere between a quarter and three quarters.

Table 2. Share of code authored by AI, by source and method.
Source Date Measure Share
GoogleApr 2026New code generated by AI and approved by engineers~75%
DX (preliminary)Q2 2026AI-authored output across participating orgs, median flat by org size51.9%
Sonar surveyJan 2026Share of all committed code42%
4.2M-developer study2026AI-written share, grew from 22% in three months26.9%
Microsoft2025 to 2026New production code in internal repositories20% to 30%

The caveat that matters most

In that same 4.2 million developer study, AI-written share rose from 22% to 26.9% in three months while measured productivity gains held around 10%. Generation is not delivery. Writing the sequencer was never the hard part of shipping a sequencer. This is not an argument against the thesis, it is the thesis: if generation compresses and coordination does not, then the value migrates to whoever owns the coordination, which is the workflow layer on top and the context layer underneath.

What that does to a build plan

Rox disclosed its own build times: a sequencer in 6 months, and enrichment and a dialer in 3 months each, on a small team with what its CTO calls a relatively large token budget. Each of those categories once carried its own company, its own funding rounds and its own multi-year roadmap. The useful way to read those numbers is not "SaaS is dead". It is that the effort required to reach parity on a known product surface has moved from a company-sized commitment to a quarter-sized one.

Table 3. Product lines by what they now cost to reach parity. Rox timelines are as disclosed; the comparison column is the historical category norm.
Product line Build time, 2026 Historically What still is not commodity
Sequencer6 monthsA funded companyDeliverability reputation, the flow it sits inside
Enrichment3 monthsA funded companyThe underlying data rights and freshness
Dialer3 monthsA funded companyTelephony compliance, carrier relationships
Scoring and prioritisationWeeksA feature teamThe event history the score is computed from
The entity and event model9 months, no outputRarely attemptedAll of it. This is the asset.

4. Why the foundation did not get cheap

A token budget buys code. It does not buy history, access, or the right to read someone's data. That is the whole asymmetry, and the numbers around it have not moved in the direction the model prices have.

  • Between 60% and 80% of the effort in an enterprise AI project goes into data preparation. Not modelling. Preparation.
  • Gartner puts the cost of poor data quality at about $12.9M a year per organisation, the most cited figure in the field, covering wasted resource, missed opportunity and operational drag.
  • B2B contact data decays at roughly 2.1% a month, about 22.5% a year on the MarketingSherpa benchmark HubSpot uses. Gartner puts general business-data decay closer to 3% a month. Behavioural and intent signals expire in weeks.
  • Gartner expects 60% of AI projects unsupported by AI-ready data to be abandoned through 2026. The blocker is the data, not the model.
  • 54% of organisations underestimate AI costs by 30% to 40%, and the underestimate is concentrated in data preparation and integration.

Put those together and the foundation reads as a maintenance problem, not a construction problem. A warehouse-native entity model with five years of event history is not something you generate. It is something you accumulate and then defend against decay, month after month, while the contact records underneath it rot at two to three percent.

Table 4. The compression test. Ask of any layer you own: what is the cost actually denominated in?
Asset Cost denominated in Compressible by tokens? Realistic time to parity
UI, CRUD, reportingEngineering hoursYes, heavilyWeeks
Integration to a documented APIEngineering hoursYesDays to weeks
A full product line at parityEngineering hours plus judgementMostly3 to 6 months
Entity resolution across 10 systemsAccess, edge cases, trustPartly6 to 12 months
Five years of event historyCalendar timeNoFive years
Permission and trust modelSecurity review cyclesNoPer customer, forever
The buyer's actual processObservation and iterationNoAs long as you serve them

5. The incumbent trap, measured

The most useful test of the thesis is not a startup. It is the largest system of record in the world, which published audited results in February 2026 and gave us an unusually clean natural experiment.

Salesforce closed FY26 with $41.5B of revenue, up 10%. In Q3 FY26 it renamed its service offerings to reference Agentforce. Its own filing is explicit about what that rename did to the underlying business: "There were no changes in the allocation of revenue between these service offerings coming from this change." Same products, same revenue allocation, new noun. Here is what the renamed lines actually grew at in Q4.

Salesforce subscription revenue growth by renamed Agentforce line, Q4 FY26 Constant currency growth rates for the quarter ended January 31 2026 against the same quarter a year earlier. Agentforce 360 Platform, Slack and Other grew 37 percent, which includes 388 million dollars of acquired Informatica revenue. Agentforce Sales grew 8 percent, Agentforce Service 7 percent, Agentforce Integration and Analytics 3 percent, and Agentforce Marketing and Commerce declined 1 percent. Constant currency growth, Q4 FY26 against Q4 FY25 includes $388M of acquired Informatica revenue Agentforce 360 Platform, Slack and Other Agentforce Sales Agentforce Service Agentforce Integration and Analytics Agentforce Marketing and Commerce 37% 8% 7% 3% -1% 0% 10% 20% 30% 40%
Figure 3. Source: Salesforce Q4 FY26 earnings release, constant currency growth by service offering. Every line carries the Agentforce name. Four of the five grow at single digits or below, and the fifth is carrying an acquisition.

None of this means Agentforce is failing. It is not. ARR reached $800M, up 169%, on more than 29,000 deals since launch, with accounts in production up nearly 50% quarter over quarter. That is a genuinely fast-growing product by any normal standard. The question is what it is growing on top of, and the scale comparison is the part that rarely gets printed.

Table 5. The arithmetic of an agent layer bolted onto an existing data model. Salesforce FY26, year ended 31 January 2026. Ratios are ours.
Figure Value What it implies
FY26 revenue$41.5B, up 10%The base is growing at the market rate
Agentforce ARR$800M, up 169%Fast, from a small base
Agentforce as a share of FY26 revenue1.9%The agent layer is a rounding error on the franchise
Agentforce plus Data 360 ARROver $2.9B, up over 200%Includes $1.1B of Informatica Cloud ARR, about 38% bought
FY26 R&D spend$5.99BAgentforce ARR equals about 13% of one year of R&D
Share of Q4 bookings from existing customersOver 60%Sold into the installed base, not winning new workflows
Records ingested by Data 360 in FY26112 trillion, up 114%The data problem is acknowledged and enormous
Of which arrived via Zero Copy53 trillion, up 310%47% of records were read where they already live

The tell

Zero Copy exists so that Salesforce can read data it does not hold, in a warehouse it does not own, without moving it. In FY26, 47% of all records Data 360 ingested arrived that way, and that mode grew at nearly three times the rate of ingestion overall (310% against 114%). The system of record is now building infrastructure whose explicit purpose is to reach the data that left it. That is not a failure. It is an admission, published in an earnings release, that the foundation layer has moved. Everything the sandwich thesis asserts about where data lives, the incumbent has already conceded in its own disclosures.

6. The build sheet

Two companies give us the two versions of the compound startup, and it is worth putting them side by side because the capital efficiency is the headline nobody says out loud.

Table 6. Two compound startups, two eras. Per-employee and per-year figures are our calculations from the disclosed totals.
Measure Rippling (SaaS-era compound) Rox (AI-era sandwich)
Shared assetThe employee recordThe warehouse-native customer graph
Products10+ lines, each above $1M ARRReplaces the work of 10+ tools
Time to a new line5 to 6 months to $1M ARR3 to 6 months to build
Headcount~5,000 (2026)Small team, large token budget
Capital deployed$450M Series G alone (May 2025)~$20M spent over 2.5 years
Implied burnn/a~$8M a year
Revenue$1B ARR (Mar 2026), up 78%Not disclosed
Revenue per employee~$200kNot disclosed
Valuation$16.8B (May 2025)$1.2B (Mar 2026)

The comparison is not apples to apples and should not be read as one: Rippling is a decade-old business at scale, Rox is a Series A-stage company with undisclosed revenue. What the comparison does isolate is the cost of getting to a multi-product position. Rippling's answer was capital and headcount. The AI-era answer, on the evidence so far, is roughly $8M a year and a deliberate refusal to ship anything for the first nine months.

Reconstructed build sequence for a sandwich-shaped compound startup A timeline over thirty months. The data foundation is built from month zero to month nine with no product output. Agents on the workflow are built from month six to month eighteen. Three product lines follow: a sequencer over six months from month fifteen, enrichment over three months from month eighteen, and a dialer over three months from month twenty one. The first visible product output arrives at month nine. Reconstructed sequencing, months from start first visible output three lines in 9 months Data foundation Agents on the workflow Product line: sequencer Product line: enrichment Product line: dialer 9 months, no product output 12 months, the workflow layer 6 months 3 months 3 months 0 6 12 18 24 30 months
Figure 4. Reconstructed from the disclosed durations (foundation about half of engineering for nine months, sequencer six months, enrichment and dialer three months each, 2.5 years total). Illustrative sequencing, not an actual project schedule. The foundation is maintained continuously after month nine; only the build phase is shown.

The shape of that chart is the whole strategy. Nine months of work with nothing to demo is not a phase you can retrofit, and it is not an engineering decision. It is a financing decision, which is the honest reason this is hard to copy: most teams cannot buy nine months of silence.

7. A test you can run on your own company

The thesis is only actionable if you can locate yourself in the sandwich. Seven questions, one point each. Answer them about the layer you actually sell.

Table 7. Where do you sit? Score one point for each yes.
# Question A yes means
1Does the freshest copy of your core entity live in a store you control?You hold the foundation
2Can an agent read five years of history in one query, without a third-party rate limit?Your context is queryable
3Would a competent team with coding agents need more than six months to reach parity with your product?You are not purely middle layer
4Is more than half your revenue tied to completed work rather than to seats?Your pricing survives the agent era
5Does the job finish inside your product, or does the user leave to finish it?You own the workflow, not a step in it
6If your UI went dark for a week, would the work still get done through you?You are above SaaS, not inside it
7Do you get paid when a customer's headcount falls?You are hedged against the seat unwind
0 to 2

You are the middle. Your product surface is now a quarter of someone else's roadmap. Go acquire a layer, in either direction, before the repricing reaches you.

3 to 5

You hold one edge of the sandwich. Decide which one you can defend and invest asymmetrically. Holding half of each is the worst position on this list.

6 to 7

You hold both edges. The product lines in between are now yours to take whenever the sales motion justifies them.

8. The demand side, in one number

Everything above is a supply-side argument about what is cheap to build. The demand-side argument is simpler. Salesforce's own 2026 State of Sales research finds sellers spend about 40% of their time actually selling, with the remaining 60% going to admin, CRM data entry, internal meetings, manual research and follow-up coordination. Roughly 68% of reps name note-taking and data input as their most time-consuming tasks, and the pattern has not meaningfully improved in five years.

Five years of feature releases did not move that number. That is the clearest available evidence that the bottleneck was never the product surface. The work sits between the tools, in coordination, and coordination is the thing the agent layer actually attacks. It is also why picking a persona and mapping the work, rather than the software the work happens inside, is the right first move.

9. Where this thesis breaks

Any argument this clean deserves its counter-case stated at full strength. Here is ours, and none of it is hypothetical.

Table 8. Failure modes, with the evidence for each.
Failure mode Evidence What it costs you
Agents demo, then never deploy79% of enterprises say they adopted agents; 11% run them in productionA workflow layer nobody uses
Projects get cancelledGartner expects over 40% of agentic AI projects to be cancelled by 2027 on unclear value, cost and governanceYour buyer's budget disappears mid-cycle
The foundation overruns60% to 80% of AI project effort is data preparation; 54% of organisations underestimate AI cost by 30% to 40%Nine months of silence becomes eighteen
Token COGS eat the margin78% of IT leaders reported unexpected charges tied to consumption or AI featuresFixed price, variable cost, negative gross margin
Distribution, not architecture, decides itAgentforce closed 29,000 deals with over 60% of Q4 bookings from existing customersYou are right and still not in the room
The incumbent buys the layer$1.1B of the $2.9B Agentforce plus Data 360 ARR is acquired Informatica CloudA rewrite you assumed was impossible arrives as an acquisition

Read together, the counter-case is not that the sandwich is wrong about the architecture. It is that architecture is necessary and not sufficient. Every failure mode above is about deployment, financing or distribution. That is a better problem to have than a structural one, but it is not a smaller one.

10. What we take from it

We build AI software, media and investing systems, so we read this as an operating instruction rather than as commentary. Five things we are applying:

  • Price the layer, not the product. Before committing to any build, run Table 4 on it. If the honest answer to "how long with coding agents" is under six months, it is a feature of your strategy, never the strategy itself.
  • Buy calendar time early. Anything denominated in accumulated history should start accumulating now, at whatever quality you can manage, because that clock does not respond to budget later.
  • Consume before you displace. Incumbent APIs are the cheapest data acquisition channel available, and using one does not commit you to competing with it.
  • Make the agent finish the job. A partial workflow leaves the user in the old tool, which keeps the old tool load bearing. The whole advantage comes from the work being done before anyone would have opened the middle layer.
  • Underwrite the silence. If the plan has nine months without visible output, that has to be financed and communicated on day one, not discovered in month five.

And the uncomfortable corollary for anyone selling a single product today: the question is no longer whether your product is good. It is whether it is compressible. Those are different questions, and only one of them has an answer that changes each quarter.

How much to trust these numbers

  • Salesforce figures are from the audited Q4 FY26 earnings release for the year ended 31 January 2026. Revenue, R&D, growth rates, ARR, record counts and the rename disclosure are quoted from that document. The ratios (1.9% of revenue, 13% of R&D, 47% of records, 38% acquired) are our arithmetic on those figures, not company disclosures.
  • Agentforce ARR and total revenue are different units. A point-in-time ARR compared against a full-year revenue figure understates the run-rate comparison somewhat. We use it as a scale indicator, not as a growth measure.
  • Rox figures ($20M, 2.5 years, 6-month sequencer, 3-month enrichment and dialer, roughly half of early engineering on the data layer) come from its CTO's own published account, 15 September 2026. They are self-reported and not independently verified. Funding and valuation figures come from investor announcements and secondary trackers.
  • Rippling ARR, growth, headcount and product-line counts come from secondary trackers and research profiles, not from audited filings. The company is private. Treat them as directional.
  • Epoch AI's price series measures the cheapest price at which a capability level is available, which is not the same as the price any given buyer pays. Epoch itself notes the fastest declines are recent and may not persist.
  • The AI-authored code percentages measure different things (committed code, approved new code, output share) under different instrumentation. The spread between 26.9% and 75% is mostly methodological and should not be read as a trend line.
  • Zylo, Gartner, Cruxy and survey-derived figures reflect their own panels and definitions. SaaS app counts in particular vary from about 106 to 342 depending on what counts as an app.
  • Figure 4 is a reconstruction for illustration. It is consistent with the disclosed durations but is not anyone's actual project plan.

Sources

Compiled September 2026. Related reading on this site: Build the Environments, Not the Workforce, which makes the adjacent argument about which layer of the AI data economy a small team should enter, and how we build these systems.

About the author

Mohinish Shaikh works on AI software, media and investing at Serverlessvc.com. The framing of this piece is owed to Shriram Sridharan of Rox and, before him, to Parker Conrad's compound startup argument. The arithmetic, the charts and any errors in either are ours.

Share this post Link copied

Prepared as market analysis. Not investment advice. Figures are dated and, in several cases, self-reported or single-source; see the reliability notes above.