The Critical-Vendor Distress Read: Stay, Hedge, Plan Exit, or Migrate Now
Score one vendor and one trigger event on five evidence-backed dimensions and reach a dated verdict that a second engineer, given the same file, would reproduce.
Key takeaways
- What the acquirer bought predicts the shutdown clock better than the price: HP bought Humane's assets and staff for $116 million and the service ended 10 days later, while Facebook bought the Parse product and gave 366 days' notice.
- A licence-and-hire deal leaves behind a company that must find a buyer; Windsurf's remainder sold to Cognition within 3 days, and without a buyer the remaining staff would have run the product alone.
- A busy changelog does not prove a vendor is safe: Parse shipped tvOS support and a dashboard redesign about a month before its shutdown was announced.
- A 12-month discontinuation clause is only as strong as its carve-outs; Google Cloud's terms allow faster exits to avoid a substantial economic or material technical burden.
- Run a full export and restore test before scoring anything, because Fauna's migration support was premium-only and Humane deleted user data 10 days after the deal.
- In Refolk's index, 258 US profiles hold third-party or vendor risk titles and the top employers are banks and health companies, so at API-first companies this read usually has no owner unless engineering takes it.
A hosted API or SaaS vendor in your production path has just announced a deal, a layoff, or a pricing change, or it has gone quiet, and you owe a decision by Friday. This guide is for the engineering manager, technical founder, DevRel lead, or technical sourcer who has to make that call. It scores five dimensions from public evidence and your own call logs and lands on one of four verdicts: Stay, Hedge, Plan exit, or Migrate now. Each verdict comes with a dated re-check, and the rubric is tight enough that a second engineer reading the same file should reach the same answer.
What does a vendor distress read decide?
It decides how much engineering time to spend moving off one commercial vendor, and how soon, after one trigger event. It does not cover open-source dependencies or contract drafting. It covers a paid or hosted vendor that has shown a distress or ownership signal.
Run the read when any of these happens:
- Ownership event: an acquisition, an asset sale, or a licence-and-hire deal is announced.
- Team event: layoffs, a notice filing, or engineering leaders leaving.
- Pricing event: a price or plan change, especially one that moves features or support behind a higher tier.
- Silence: the changelog stops, SDK releases stop, or issue responses slow down.
The read ends in one of four verdicts. I use the same four every time so that reads stay comparable across vendors and quarters.
| Verdict | What you do | Default re-check |
|---|---|---|
| Stay | Keep building on the vendor as planned | Next trigger, or one quarter |
| Hedge | Add an abstraction layer and a tested fallback | Monthly |
| Plan exit | Schedule a dated migration on the roadmap | Every two weeks |
| Migrate now | Pause roadmap work and move | Weekly until cutover |
An abstraction layer is a single internal interface that your code calls, with the vendor's SDK hidden behind it, so a replacement only has to satisfy that interface. A fallback is a second implementation behind that interface that you have actually run against production-shaped traffic.
Why does what was bought set the shutdown clock?
What the acquirer actually bought predicts how long the service survives better than the price does. An acquirer that buys the product inherits customers it wants to keep. An acquirer that buys IP and staff has no reason to keep the service running.
The four documented cases show the spread:
| Case | Deal type | Announcement to end | Export offered |
|---|---|---|---|
| Humane | Assets and staff to HP | 10 days | Self-download, then deletion |
| Fauna | No deal, capital ran out | 70 days | JSON snapshots to S3 |
| Parse | Product acquisition | 366 to 368 days | Migration tool and open source |
| Windsurf | Licence and hire | Product continued | Not applicable |
The detail behind each row matters more than the number:
- Humane. On 18 February 2025 HP agreed to buy most of Humane's assets for $116 million. HP took the CosmOS AI platform, Humane's technical staff, and access to over 300 patents and patent applications. It did not buy the product line. Sales stopped that day, the service shut down on 28 February 2025, and Humane deleted user data stored on its servers after 12 PM PST that day. There was no export tool.
- Fauna. There was no buyer. The board and investors decided it was not possible to raise the capital needed to keep going on its own. The shutdown was announced on 21 March 2025 and the service was switched off on 30 May 2025, after which all accounts were deleted. Before the cutoff Fauna ran a brownout, meaning a deliberate period where all queries returned a 429 response code, to force stragglers to notice.
- Parse. Facebook bought Parse in 2013 for $85 million and ran it for about three years. In 2014 it powered 500,000 apps. On 28 January 2016 Facebook announced it would shut down on 28 January 2017, and it actually ran until 30 January 2017. Facebook provided a migration tool and open-sourced the server.
- Windsurf. Google paid $2.4 billion for a licence plus compensation and hired the leaders. The remaining company sold its assets to Cognition three days later, and the product continued.
The common shorthand for the short-clock deals is acquihire, a deal where the acquirer mainly wants the team. Split it in two, because the two halves behave differently. In an assets-plus-staff deal the product is left with no owner. In a licence-plus-hire deal a remainder company still owns the product but has lost its leaders and has to find a buyer. Windsurf's remainder found one in three days. If it had not, the remaining staff would have been running the product alone.
What the acquirer bought predicts the shutdown clock better than what the acquirer paid.
How do you score the five dimensions?
Score each dimension 0, 1, or 2 against the observable evidence described below, where 0 is reassuring and 2 is alarming. Add dimensions 1 to 4 to get distress (0 to 8). Keep dimension 5 separate as exposure, because it describes your system rather than the vendor.
| Dimension | 0 | 1 | 2 |
|---|---|---|---|
| 1. Deal type | No deal, or product bought with its customers | Licence and hire, or deal not yet classified | Assets or IP and staff bought, product not |
| 2. Investment cadence | At own baseline, several authors | Below baseline, or fixes only | SDK releases stopped, issues unanswered |
| 3. Team and runway | Stable in two sources | Drop in one source, or leaders leaving | Drop in two sources, WARN filing, or failed raise |
| 4. Exit protection | Narrow clause, restore passed | Broad carve-outs, or export untested | No clause, or export fails or is tier-gated |
| 5. Blast radius | Off the vendor in under 10 days | 10 to 70 days | Over 70 days |
When the evidence for a dimension is missing, score the worse option and write "unknown" next to the score. A missing fact is not a reassuring one.
Dimension 1: deal type and acquirer sunset record
What it proves: whether anyone on the buying side owns the product and its customers. Read the acquirer's own press release, not the coverage, and list what changed hands: product, customers, IP, staff, or a licence. Then check the acquirer's published deprecation policy and what it did with products it bought before.
When it lies: headlines say "acquired" when the product was not acquired. HP's purchase of Humane was widely reported as an acquisition, but HP took the assets and staff, not the product line. One announcement can also split in two: the Windsurf deal became a licence to one buyer and an asset sale to another within three days. Re-classify when the second shoe drops.
Dimension 2: product investment cadence
What it proves: someone is still assigned and paid to ship this product. Use three sources: the public changelog, release tags in the vendor's SDK repos, and time to first response on its public issue tracker. Compare each against the vendor's own baseline over an earlier window of the same length, not against other vendors.
When it lies: shipping can continue right up to the announcement. One Parse user noted new features such as tvOS support and a dashboard redesign that had rolled out about a month before the shutdown post. Count who is shipping as well as how much. If SDK release authors shrink to one or two names, or founders' commits stop while a handful of remaining staff keep tagging releases, score 1 even if the volume looks healthy.
Dimension 3: team and runway signals
What it proves: whether the people who run the service will still be there in a quarter. The inputs are the engineering headcount trend, departures of engineering leaders, layoff filings, and any statement about funding.
When it lies: three ways.
- Single-source headcount. For Windsurf, one outlet reported a 250-person team and another source gave 112 people for the same month, a 2.2x gap. Never score this dimension from one number.
- WARN filings. The US WARN Act covers employers with 100 or more employees and requires 60 days' notice. Most API startups are below that size, so no filing proves nothing.
- Runway arithmetic. No published vendor-risk framework states a runway cutoff, so do not invent one. The strongest documented runway signal is a plain statement like Fauna's, that the capital to continue could not be raised.
For a second headcount source that does not depend on press figures, count the people who currently list the vendor as their employer. In Refolk you can ask for current engineers at the vendor in plain English, re-run the same request on the re-check date, and compare the two lists.
Dimension 4: contractual exit protection
What it proves: how much warning and help you are legally owed, and whether you can actually take your data out. Pull the discontinuation clause, every carve-out, the export and migration terms for your plan tier, and any escrow terms.
Published minimums give you something to compare your contract against:
| Provider | Minimum notice | Carve-out or caveat |
|---|---|---|
| Google Cloud (GA services) | 12 months | Law, security, economic burden |
| Google Maps Legacy | 12 months | None noted |
| Google Ads API | Months | Set per version |
| Salesforce API | 1 year notice, 3 years support | None noted |
When it lies: the headline period is the least important line in the clause. Google Cloud's 12-month commitment carves out discontinuations needed to comply with law, to address a material security risk, or to avoid a substantial economic or material technical burden. The last carve-out is broad enough to cover a failing product. Google Cloud's liability cap for free services is $5,000, so a free-tier dependency has almost no contractual recovery.
Plan tier matters as much as the clause. Fauna offered migration support only on premium plans, so smaller customers had 70 days to manage alone.
Escrow, an arrangement in which a third party holds a deployable copy of the service and releases it on a defined trigger, has one documented rescue. An airline's SaaS vendor went into administration only a few months into the service, and the escrow agent brought a dormant cloud replica into live status. That worked because the replica had been built and checked before the failure. Release triggers that need a court finding can delay access by months, which is longer than two of the four documented shutdown windows.
What actually protects you on exit day
- Headline notice periodThe 12-month or 1-year figure everyone quotes
- Carve-outsLaw, security, and economic-burden exits that shorten the period
- Plan-tier termsWhether export and migration help apply to your tier
- Verified escrowA deposit someone has checked, with a fast release trigger
- Passed restore testYour data running somewhere the vendor does not control
How do you measure blast radius and switching cost?
Blast radius is the part of your product that breaks when the vendor stops answering. Switching cost is the calendar time to remove the vendor: engineer-weeks divided by the engineers you can actually staff, plus your client release lead time. Score dimension 5 on calendar time, because the vendor's clock runs in days, not effort.
Measure from your own systems:
- Call sites: search the codebase for the SDK import and the vendor's API hostnames. Count files and services, not lines.
- Monthly calls: take them from your logs or the vendor's billing page. These tell you which call sites carry real traffic.
- Data held at the vendor: the size and shape of what must be exported, and whether the export includes everything your restore needs.
- Client release lead time: server-side changes ship when you deploy, but mobile clients need a shipped build in users' hands before the cutoff. Parse migrators learned this the hard way.
- Staffable engineers: the people you can really free up this month, not total headcount.
Then score against the two short documented clocks. Under 10 days scores 0, because you could survive a Humane-style exit. Between 10 and 70 days scores 1, because you could survive a Fauna-style exit if you started at once. Over 70 days scores 2.
The weakest input is the engineer-week estimate. No published threshold exists, and the only practitioner figure I can find is a low-end Parse claim that you could have a Parse server running "with 20 minutes of your time and only $5/month". Treat that as the floor for a trivial app, not an estimate for yours. The better calibration is a 30-minute call with someone who has already done the same migration at production scale. That is a people-search problem, and it is the kind of search Refolk handles in one sentence.
How do the scores become a verdict?
Place the vendor on two axes, distress (dimensions 1 to 4) and exposure (dimension 5), then apply one override: if you need longer to get off the vendor than the shortest documented clock for this deal type, the verdict is Migrate now. Distress is high if dimensions 1 to 4 sum to 4 or more, or if dimension 1 alone scores 2. Exposure is high if dimension 5 scores 2.
The cut at 4 is my convention, not a published threshold, because none exists. Keep it fixed across every read so that two verdicts a quarter apart are comparable.
Distress against exposure
Use the clock that matches the deal label from dimension 1:
- Assets or IP plus staff: 10 days, the Humane case.
- Licence plus hire: 10 days until a buyer for the remainder is named, then re-score on the buyer's own deal type.
- No deal, capital exhausted: 70 days, the Fauna case.
- Product acquisition: the contract's notice period, read with its carve-outs. Parse's 366 days is the generous end, not a promise.
- No deal and no capital signal: the contract's notice period.
Write the read down in one file per vendor per trigger. The point of the file is that a second engineer can score it without talking to you.
Vendor: Trigger event and date: Scorer: Second scorer: EXPORT (done first) Last full export - date, size: Restore target - result (pass/fail), date: Migration support on current plan tier - yes/no, source: D1 Deal type (0/1/2): What was bought (quote the acquirer's release): Label: product acquisition / assets or IP + staff / licence + hire / no deal Acquirer's published deprecation policy: D2 Investment cadence (0/1/2): Changelog entries - current window vs prior window: SDK release tags - current vs prior, and authors: Issue first-response time - current vs prior: D3 Team and runway (0/1/2): Headcount source 1: Source 2: WARN filing (n/a if under 100 employees): Engineering leader departures: Funding statements: D4 Exit protection (0/1/2): Notice clause (quote): Carve-outs (quote each): Escrow - deposit verified on (date) / none: D5 Blast radius (0/1/2): Call sites: Monthly calls: Data held at vendor: Engineer-weeks: Engineers staffable: Client release lead time: Calendar time to off the vendor: Distress (D1+D2+D3+D4): Exposure (D5): Clock for this deal type: Override applied - yes/no, which: VERDICT: Stay / Hedge / Plan exit / Migrate now Re-check date: Event that would move the verdict early:
One file per vendor per trigger. Quote sources rather than paraphrasing, and write "unknown" rather than leaving a line blank.
The procedure: from trigger to a dated verdict by Friday
The procedure runs export first, then classification, contract, signals, and your own exposure, and ends with a blind second score. Export comes first because the shortest documented window was 10 days and ended in data deletion. Nothing else in the read can bring that data back.
Trigger to dated verdict
- Start the export and restore testKick off a full export of everything the vendor holds and restore it into an environment the vendor does not control. Record date, size, and result in the file.
- Classify the deal from what was boughtRead the acquirer's own announcement, list what changed hands, and give the deal exactly one label.
- Pull the discontinuation clause and carve-outsQuote the notice period, every carve-out, plan-tier export and migration terms, and any escrow release trigger.
- Read the outside-in signalsCompare changelog, SDK release, and issue-response cadence against the vendor's own baseline, check release authors, and triangulate headcount from two sources.
- Measure blast radius and switching timeCount call sites, monthly calls, and data held, convert engineer-weeks to calendar time with real staffing, and add client release lead time.
- Score, apply the override, place on the matrixSum dimensions 1 to 4 into distress, read dimension 5 as exposure, apply the clock override, and write one verdict.
- Second-score blind and date the re-checkA second engineer scores the same file without seeing your numbers, differences are settled by evidence, and a re-check date before the shortest plausible clock goes in the file.
The steps fit a working week if the export starts on the first day. Steps 2 and 3 can run in parallel with legal. Step 4 is mostly reading. Step 5 is the slowest, because it needs a code search and a staffing conversation.
For step 3, the contract often will not answer every question. Ask the vendor in writing and file the reply.
Please confirm in writing, for this account and its current plan tier: 1. The discontinuation notice period that applies to the services this account uses, and the clause it comes from. 2. Every condition under which that notice period can be shortened. 3. Whether a full data export is available on this plan tier, in what format, and what it excludes. 4. Whether migration assistance is offered on this plan tier if a service is discontinued. 5. Whether the recently announced ownership or organisational change alters any term of this agreement, and if so which. 6. Who the named contact for this account will be over the next two quarters.
Send by email, not on a call. Any question left unanswered by the re-check date is scored as the worse option.
How does the distress read go wrong?
Most bad verdicts come from reading a signal as proof when it can be produced by a healthy vendor and a dying one alike. There are seven documented ways this goes wrong. Each one comes with the check that catches it.
- A busy changelog read as safe. Parse shipped tvOS support and a dashboard redesign about a month before its shutdown was announced. Volume shows work already in the pipeline, not a commitment to next year. Check: look at who authors SDK releases and changelog entries, and whether that group is shrinking.
- Headcount trend from one source. Two sources for Windsurf's team differed by 2.2x. A single figure can look like growth or collapse depending on which one you pick. Check: use at least two independent sources and score the trend only where they agree.
- No WARN filing read as no layoffs. WARN only applies to employers with 100 or more employees. Most API startups fall below that, so silence in the filings tells you nothing. Check: for vendors under that size, ignore WARN and rely on the two-source headcount trend.
- Trusting a 12-month clause. Google Cloud's terms allow faster discontinuation to avoid a substantial economic or material technical burden. A failing product is a textbook economic burden. Check: quote every carve-out in the file and score dimension 4 on the broadest one.
- Assuming escrow works. Release triggers that require a court finding can delay access by months. The one documented rescue worked because the replica had been checked before the vendor failed. Check: confirm when the deposit was last verified and whether the trigger needs a court.
- Assuming export still works. Fauna's migration support was premium-only, and Humane offered no export tool at all. An export that worked last year, or on another tier, proves nothing today. Check: run a full export and restore before scoring, and date the result.
- Treating "acquired" as "the product was acquired". HP's deal was reported as an acquisition of Humane, but HP bought assets, platform, staff, and patents, not the product line. Check: label the deal from the acquirer's own list of what it bought, never from a headline.
One structural failure sits behind all seven: the read happens once, at the trigger, and is never repeated. A vendor that scored Stay in one quarter can be a 10-day clock in the next. The re-check date is part of the verdict, not an afterthought.
Who owns this read, and how do you keep it current?
At most API-first companies nobody owns this read, so an engineering manager or technical founder has to take it on by name. Formal vendor-risk work sits in regulated industries, where a regulator requires vendor review. Engineering teams have no such requirement.
| Segment | Country | Profiles in Refolk's index | Top title share (25 sampled) |
|---|---|---|---|
| Vendor or third-party risk | US | 258 | Vendor Risk Analyst, 36% |
| Vendor or third-party risk | UK | 65 | Third Party Risk Manager, 72% |
| Developer Relations | US | 285 | Developer Advocate, 64% |
Two ratios stand out. The US vendor-risk count is 3.97x the UK count. And there are slightly more US DevRel profiles than vendor-risk profiles (1.10x), which suggests the people closest to vendor APIs in developer tooling are advocates rather than risk analysts. The mechanism is the same in both cases: vendor review lives where a regulator requires it. If you want to borrow the method from people who do formal vendor review for a living, Refolk can find third party risk managers at UK banks or vendor risk analysts at large US lenders, and an hour with one of them is a cheap check on your rubric.
Keep the read current with four habits:
- Re-score on every new trigger, not only on the scheduled date. A second deal, a brownout, or a tier change resets the read.
- Re-run the restore test on each re-check date. Export formats and tier rules change without notice.
- Diff the terms. Providers publish versioned, dated copies of their terms. Save the version you scored against and compare the live one at each re-check and at every renewal.
- Keep the cut fixed. Do not move the distress cut of 4 to fit a verdict you already wanted. Change it only between reads, and write down why.
Before you send the verdict
- A full export completed and a restore into an environment the vendor does not control passed, with the date recorded.
- The deal carries one label, taken from the acquirer's own list of what it bought.
- The discontinuation clause and every carve-out are quoted word for word in the file.
- Plan-tier export and migration terms are confirmed for the tier you actually pay for.
- The headcount trend is scored from at least two independent sources.
- A missing WARN filing was not counted as evidence for a vendor under 100 employees.
- SDK release and changelog authors were checked, not only counts.
- Switching time is in calendar days and includes client release lead time.
- The clock override was applied and its result is written in the file.
- A second engineer scored the file blind and reached the same verdict.
- The re-check date falls before the shortest plausible clock for this deal type.
Questions practitioners ask
Our API vendor was acquired. Should we migrate?
It depends on what the acquirer bought, not on the headline. If the acquirer bought the product and its customers, as Facebook did with Parse, the service may run for years and give long notice. If it bought assets, IP, and staff, as HP did with Humane, the service can end within days. Read the acquirer's announcement, label the deal, run an export and restore test today, then score the other four dimensions before choosing a verdict.
How long is a typical vendor sunset notice period?
No median has been published, so treat any 'typical' figure with suspicion. Published minimums include 12 months for Google Cloud GA services and one year of notice for Salesforce API versions, while Google Ads API windows are often measured in months. Documented real cases range from 10 days (Humane) through 70 days (Fauna) to 366 days (Parse). Read your own contract's carve-outs, because they decide whether the headline period actually binds.
What are the signs a SaaS vendor is going out of business?
The useful signals are a slowdown against the vendor's own changelog and SDK release baseline, slower issue responses, a headcount decline confirmed by two independent sources, leadership departures, and statements about funding. Each can mislead. Parse shipped features a month before its shutdown announcement, headcount sources for Windsurf differed by 2.2x, and WARN filings only cover employers with 100 or more staff. Score them together, never alone.
Does an acquihire mean the API will shut down?
Not always, but it shortens the clock sharply. When an acquirer takes assets and staff without the product, nobody has a reason to keep the service up: Humane ended 10 days after HP's deal. In a licence-plus-hire deal the remaining company has to find a buyer. Windsurf found Cognition within 3 days and the product continued. Until a buyer is named, treat a licence-plus-hire vendor as if it were on the short clock.
Is SaaS escrow worth paying for?
Only if the deposit is verified and the release trigger is fast. In one documented case an airline's SaaS vendor entered administration a few months into the service, and the escrow agent brought a dormant cloud replica into live status. That worked because the replica already existed and had been checked. Release triggers that need court verification can delay access by months, which is longer than some real shutdown windows.
How do I estimate the switching cost of a critical API?
Measure it as calendar time, not effort alone. Count call sites and monthly calls, size the data held at the vendor, estimate engineer-weeks, divide by the engineers you can actually free up, then add client release lead time, because mobile apps need a shipped build before the cutoff. No published threshold exists, so calibrate the estimate by talking to engineers who have done the same migration.
Try it on the search you came here for
Stop building boolean strings. Just describe the person.
Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.
01Describe them
One plain sentence. Role, city, stack, stage, whatever matters to you.
02I read the web live
GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.
03You read the shortlist
Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.
- Staff backend engineers in NYC who shipped Rust in production
- Series A fintechs in SF under 50 people, growing headcount this year
- Maintainers of fast-growing Rust web frameworks on GitHub
- No boolean, no filters, no seat to buy. One box.
- Read at search time, so a profile updated yesterday counts today.
- Every step visible as it runs, every name with its reason.
500 free credits on sign-up. No card, no demo call. See real searches.