Innodata Inc [INOD] — Research

As of 2026-07-29 · framework v1.5.1 · Tier-2 memo · spot $57.56 (2026-07-28 close, Alpaca SIP)

This memo issues no position verdict. It scores Criteria and outputs an analysis. The book decides.

0. Headline answer to the brief's trap

The brief asked whether a 70% CAGR built on two or three customers on renewable statements of work is a different asset from 70% recurring growth. It is, and the filings say so explicitly. But the trap does not land where it was aimed:

So: revenue is uncommitted, cancellable on 30–90 days, and 56% from one counterparty — but it is organic, cash-collected, and clean. This is not an Applied Optoelectronics or 10x Genomics accounting problem. It is a counterparty-concentration and contract-tenure problem, which is a different, and more honest, risk.

The brief's second warning — "the terminal-margin assumption for a services business should be far below a software one" — turns out to be aimed at the wrong parameter. The screen's terminal margin (15.8%) is already a services margin and INOD already earns it. The software-like assumption is hiding in the exit multiple: 27.2x EBIT, drawn from a 72-name universe-wide growth-matched set, against a services comp set that trades at 6.8–9.0x. See INOD_Valuation.md §3.

1. What the company is

Innodata (Nasdaq: INOD; CIK 0000903651; SIC 7374) is a 36-year-old data-engineering services business that pivoted into AI training data. It supplies AI training and post-training data, model evaluation, alignment and safety work, and AI model deployment/integration to frontier-model developers, delivered by a large offshore expert workforce in the Philippines, India, Sri Lanka, Canada and Israel.

The customer base "includes five of the companies commonly referred to as the 'Magnificent Seven', as well as several leading artificial intelligence research labs and model developers" (10-K, Item 1). Innodata has never named the top customer in any filing, and this memo does not name it either — no filing, transcript or press release identifies it, and any attribution would be fabrication.

Segments. Historically three: Digital Data Solutions (DDS), Synodex (medical-records structuring) and Agility (media-intelligence PR software). Effective Q1 2026 the company collapsed to a single reporting segment. That is a disclosure degradation and is scored as one — see §4.

FY2025 segment split (final year disclosed):

Note the Synodex line. The company has an in-house, disclosed precedent of a customer contract termination reducing a segment's revenue. That is the bear case's named cause, already observed at small scale.

2. The mechanism — named, evidenced, dated

Not a narrative. Four specific, filed items:

(a) A named second Big Tech program worth ~$51m in 2026. Press release, 2026-05-07 (8-K EX-99.1):

"We are also announcing today a new set of engagements with one of the world's leading Big Tech companies that we expect could generate approximately $51 million of revenue this year. Twelve months ago, our revenue from this customer was zero; this year we expect it to become our second-largest customer."

This is verifiable against the concentration disclosure and it checks out: Q1 2026 had a second customer at 17% of revenue = $15.3m, where no second customer reached 10% in Q1 2025. So ~$15m of the $51m is already recognised and ~$36m is scheduled across Q2–Q4 2026 — roughly $12m per quarter of already-identified incremental revenue. This is the single most load-bearing fact in the growth case, and it is disclosed rather than inferred.

(b) Broadening beyond the anchor account. Same release: "In Q1, revenue from our other Big Tech customers, in the aggregate, grew 453% year-over-year." Cross-checkable: the top customer fell from 61% → 56% of revenue while total revenue grew 54%, i.e. the non-top-customer base grew from $22.8m to $39.6m = +74%. The 453% figure is a narrower cut (Big-Tech-only) than the residual base, so it is not directly reconcilable from filings — treated as unverified management arithmetic, not adopted.

(c) A named new product. The Evaluation and Observability Platform, launched in beta in Q1 2026, described as "a control plane for agentic systems." First platform engagement closed at $1m of revenue with a hyperscaler; 15 additional companies actively evaluating; discussions with two hyperscalers about channel partnerships. First platform revenue is immaterial today; the significance is directional (see §5).

(d) A raised, then reaffirmed, full-year guide. FY2026 revenue growth guidance raised from ~35%+ to ~40%+ on 2026-05-07, then reaffirmed on 2026-06-17 in a standalone 8-K issued mid-quarter — thirteen days after the stock's all-time-high close and eight sessions into a −40% drawdown. Management called the guide "prudent" and said "there are several potentially large programs we have not yet included in our forecast."

The arithmetic on that guide is worth stating plainly: ~40% growth on FY2025's $251.7m is $352.3m, but Q1 2026 annualised is $360.4m. The raised guide therefore implies zero sequential revenue growth for the remaining three quarters of 2026 — while $36m of an already-named program is still to be recognised. Either the guide is materially conservative, or Q1 contained something that does not repeat. Nothing in the filings indicates the latter (no one-off revenue, no settlement, no catch-up). This is the most actionable observation in the memo and it is the basis of the 12-month target.

3. Accounting quality — is the reported growth real?

Verdict: yes, with two distortions that must be stripped out and one item I could not resolve.

3.1 Growth is organic and cash-backed

• Zero acquisitions. FY2024, FY2025 and Q1 2026 investing cash flow is capex only ($7.7m, $11.1m, $2.4m). Goodwill is flat at $2.0m. Every dollar of the 48%/54% growth is organic.
• Cash conversion. FY2025 CFO $46.8m on net income $32.2m = 1.45x. FY2024 $34.9m on $28.7m = 1.22x. FY2025 FCF = $35.6m; TTM FCF ≈ $62.0m.
• Sloan accruals FY2025 ≈ −10.1% of average assets. Negative accruals = high earnings quality.
• DSO is not being stretched — the classic receivable-stuffing tell is absent:

Receivables fell $0.6m in the quarter revenue rose 24.5% sequentially. That is the opposite of the AAOI pattern the brief flagged. It is a genuinely strong signal.

3.2 Distortion 1 — FY2024 EPS is tax-benefit-inflated. Do not use it as a base.

FY2024 reported net income $28.7m on operating income of $24.3m — net income exceeded operating income. The effective tax rate was negative 17.1%, per the 10-K reconciliation: change in valuation allowance (30.7) and effect of stock-based compensation (64.8). The 10-K states directly that Q3 2024's $17.4m included "a favorable one-time valuation allowance adjustment of $6.0 million reflecting the release of a reserve related to accumulated net loss carry forward."

Consequence: reported EPS went $0.89 (FY24) → $0.92 (FY25), apparently +3%. Pre-tax income went ~$24.5m → ~$42m, +71%. Anyone anchoring on the EPS line will conclude earnings growth has stalled; that is a tax artifact. Anyone anchoring on FY2024 EPS as a clean base is wrong in the other direction. Use pre-tax income or operating income.

3.3 Distortion 2 — Q1 2026 operating cash flow is working-capital-flattered

Q1 2026 CFO of $37.3m on $90.1m of revenue (41% CFO margin) is not repeatable. It includes: - accounts payable / accrued expenses / other liabilities +$11.6m - accrued salaries, wages and benefits +$4.9m

i.e. ~$16.5m of the $37.3m is an accrual build. Underlying cash generation ex-working-capital is ~$23m, which reconciles to Adjusted EBITDA of $25.0m less cash taxes. The headline CFO number should be halved before annualising.

3.4 Unresolved item — the accrued-expense jump

Accrued expenses and other liabilities went $9.6m (2025-12-31) → $27.2m (2026-03-31), +183% in one quarter against +24.5% revenue. That is 19% of quarterly revenue appearing in one line. The 10-Q provides no component breakdown and I could not resolve it from the filings. Most plausible benign reading is accrued offshore delivery/contractor cost and bonus accrual on a fast headcount ramp; that is an inference, not evidence. This is the one open accounting question on the name and it should be the first item on the Q2 call.

3.5 Stock-based compensation is real, growing faster than revenue, and excluded from every non-GAAP metric

Q1 2026 SBC annualises to ~$24m. Adjusted EBITDA of $25.0m is struck after adding back $5.9m of SBC — i.e. 24% of the headline Adjusted EBITDA is added-back stock compensation. GAAP operating income for the quarter was $16.9m. This is disclosed and reconciled in the release (no Twist-style concealment), but the $25.0m figure is the one in every headline and it is not a cash-economics number.

The dilution is showing up: diluted WASO 35.572m vs 32.655m shares outstanding — a 8.9% spread, and shareholders approved an Amended and Restated Equity Compensation Plan at the 2026-06-04 annual meeting.

3.6 Disclosure degradation — flagged

Effective Q1 2026 Innodata retired three-segment reporting (DDS / Synodex / Agility) for a single segment. The stated reason is "the transformation of the Company's business strategy and operating model." The effect is that the declining businesses (Synodex −8%, Agility +9%) are no longer separable from the growing one. Customer-concentration disclosure was retained and, to the company's credit, improved — Q1 2026 disclosed a second 10%+ customer that it was not obliged to break out. Net: one disclosure lost, one gained. Not a Twist pattern, but the segment loss removes the ability to verify DDS growth independently going forward.

4. Governance and insider behaviour — the largest single finding after concentration

Insiders sold $158.3 million of stock in a five-week window (2026-05-12 to 2026-06-16), at an average price of ~$95–99. The stock is $57.56 today — 41% below their average execution.

All figures parsed directly from Form 4 XML, non-derivative transaction code S, disposals only, 2026 YTD:

Context: 1.62m shares ≈ 5.0% of shares outstanding, and $158.3m ≈ 8.4% of today's $1.88bn market cap, sold in five weeks by five insiders including the CEO, the Interim CFO and two directors.

Timing relative to disclosure: selling began five days after the 2026-05-07 release that produced an +86.0% single-day move, and ended twelve days after the 2026-06-04 all-time-high close of $121.50.

On Rule 10b5-1: none of the Form 4s filed in the window carries the Rule 10b5-1(c) affirmative-defense checkbox (aff10b5One = 0 on all 20 filings 2026-05-01 onward), and none of the 21 Form 144s filed in the window reports a plan adoption date. I am stating what the filings do and do not contain; I am not asserting the sales were not planned.

Other governance items, stated without editorialising: - The company operated with an Interim CFO (Espineli) through its highest-growth period; a permanent CFO, Jayant Chauhan, was announced 2026-06-17 effective 2026-07-06, arriving from SVP of Mergers & Acquisitions at Mphasis. An M&A-background CFO at a company with $117m of undeployed cash and a stated interest in "possible business acquisitions" is a signal worth tracking. - Independent auditor is BDO India Services Private Limited (ratified 2026-06-04) — the India member firm, an uncommon choice for a US registrant of this size. - Say-on-pay passed with 10.15m for / 0.37m against, but on a base where 8.88m broker non-votes exceeded the total against on every proposal.

5. Transcript mention-frequency (required core metric)

Substrate substitution, declared: transcripts are not on EDGAR and no transcript corpus was reachable in this run. I ran the metric over the 14 quarterly earnings press releases (8-K EX-99.1), Q4 2022 → Q1 2026, which contain the CEO's prepared quotes. This is a proxy, not the specified corpus, and it is weaker because release length collapsed from ~1,950 to ~900 words in mid-2024. All counts below are normalised per 1,000 words, per the reference file's explicit warning about the ISRG length artifact.

Run generatively — before forming a view — the anomalies are:

Emerging from zero:

Decaying to zero:

Open questions this generates, recorded with no interpretation attached (per the reference file's step 2):

1. Why did observability appear for the first time in 14 quarters in 2026Q1? — Resolved: it names the Evaluation and Observability Platform beta. Corroborated in the same release.
2. Why did sovereign and federal spike together in 2025Q3 and then vanish? — Unresolved. A stated Federal Practice exists in the 10-K, but no revenue, contract or award is disclosed for it. If a federal pipeline had converted, this is where it would have shown. It did not. Recorded as a decayed initiative.
3. Why has annotation disappeared completely while revenue tripled? — Consistent with the reported mix shift from labelling toward evaluation/alignment/deployment, which carries the gross-margin expansion (40%→44% GAAP). Not independently corroborated — the segment retirement removed the disclosure that would have tested it.
4. hyperscaler re-entering after ten silent quarters coincides with the first platform sale and two named channel discussions.

Provenance note: items (a), (c) and (d) of the mechanism in §2 were sourced from the filings before this run. Item (2) above — the decayed sovereign/federal initiative — is the only finding this metric produced generatively, and it is a negative one. Logged as such.

6. Screen validation — three discrepancies found, none silently adopted

Per the brief: share count, TTM revenue and net cash verified against primary filings.

Two further screen defects, both material to the ranking:

(i) Basic shares understate EV by 9%. With diluted WASO of 35.572m (a 3.0m spread driven by SBC), market cap is $2,047m not $1,879m and EV is $1,933m not $1,767m. EV/Sales is 6.82x, not 6.23x. This flows straight into the implied path.

(ii) The screen mixes periods: TTM revenue with a last-fiscal-year margin. op_margin_pct 15.8 is FY2025's margin (39,873 ÷ 251,663), applied to TTM revenue to give ev_ebit 39.3. TTM operating income is actually $48.5m (17.1%) — Q2'25 8,911 + Q3'25 11,759 + Q4'25 10,931 + Q1'26 16,900. Q1 2026 alone ran 18.8%. Corrected EV/EBIT on diluted shares is 39.8x — close to the screen's number only because the two errors offset.

Price series: verified real. The brief flagged 122% vol as near the data-quality ceiling. Independently pulled 894 daily SIP bars, 2023-01-03 → 2026-07-28. My realized vol: 252d 105.8%, 126d 125.8%, 63d 156.5%, 21d 57.6%. The screen's 121.5% is within that family. The vol is genuine and event-driven, not a data artifact: 2026-05-08 +86.0% in one session on Q1 earnings; peak close $121.50 on 2026-06-04; today $57.56 = −52.6% from the peak in eight weeks. Every large move maps to a dated, filed event.

7. Criteria scorecard

Archetype declared: COMPOUNDER (already profitable). Noted: margins are expanding, not stable (14.2% → 18.8% op margin over five quarters), so the INFLECTION standard is scored in parallel and it also passes.

8. Downside case — named cause, MEASURED

The named cause is not "AI spending slows." It is: the customer representing 56% of revenue and 65% of receivables reduces, delays or terminates its statements of work, which it may do on 30–90 days' notice without penalty.

Three ways that happens, in descending likelihood: 1. Programme completion without replacement. The 10-K risk language is explicit: "potential inability to replace projects that are completed, canceled or reduced." This is the base mechanism. 2. In-sourcing. A frontier lab builds internal data operations, or substitutes synthetic data for human-expert annotation. No filing evidence either way; a structural, not company-specific, risk. 3. Price concession. 10-K: customers "may also request that we modify certain key terms of our agreements with them as a condition of continuing to do business with us."

In-house precedent this is not hypothetical: Synodex revenue fell 8% in FY2025, the 10-K attributing it "primarily to termination of a customer contract."

No going-concern case is argued. Net cash of $114.8m, no drawn debt, a $50m undrawn revolver and positive FCF at every scenario revenue level mean solvency is not the risk. The risk is a 50–70% equity drawdown, not a zero. Probabilities above are analyst judgment, logged for Brier scoring.

These bear scenarios are not remote-tail constructions: the stock already fell 52.6% in eight weeks in June–July 2026 on no disclosed negative news — guidance was reaffirmed on 2026-06-17 inside that drawdown. The name trades a −50% move on sentiment alone.

9. What is unsupported — stated explicitly

Per the brief's "never fabricate" rule, everything below is not established by this memo:

1. The identity of the 56% customer. Not named in any filing. Any attribution would be fabrication.
2. The identity of the second (17%, ~$51m) customer. Described only as "one of the world's leading Big Tech companies."
3. The "453% growth in other Big Tech customers" figure — management arithmetic on a cut not disclosed in the financials. Not adopted. The verifiable analogue is the residual base at +74%.
4. The $17.2m single-quarter jump in accrued expenses and other liabilities. No breakdown filed. Unresolved.
5. Any FY2027 estimate. No FY2027 guidance exists. The 12-month target's NTM band is constructed from FY2026 guidance plus the disclosed $51m programme schedule, and is labelled as such — it is not consensus.
6. Street consensus, target prices and analyst counts. Alpha Vantage quota exhausted. No external target is quoted anywhere in this memo, so the valuation.md "sanity band" check against a professional target could not be performed. Recorded as a gap.
7. Whether the insider sales were made under pre-existing Rule 10b5-1 plans. The filings carry no such checkbox and no plan adoption date; that is all that can be said.
8. Federal / sovereign AI revenue. A Federal Practice is described in the 10-K. No contract, award or dollar of revenue is disclosed. Mention-frequency shows the theme spiked in 2025Q3 and vanished.
9. Any Q2 2026 earnings date. Not announced in any filing as of 2026-07-29. The catalyst calendar carries the historical analogues and labels the inference.
10. Terminal-year growth. If INOD is still compounding above ~20% in 2031, no services comparator brackets it and the exit multiple is UNIDENTIFIED, not 9.0x and not 27.2x. This is declared in the valuation file rather than papered over.

Sources

All primary. No secondary or press sources used.

• 10-K FY2025, filed 2026-02-26, accession 0001104659-26-020655
• 10-Q Q1 2026, filed 2026-05-07, accession 0001104659-26-057270
• 10-K FY2024, filed 2025-02-24, accession 0001410578-25-000194
• 8-K + EX-99.1, 2026-05-07 (Q1 2026 results), accession 0001104659-26-057150
• 8-K, 2026-06-17 (guidance reaffirmation; CFO appointment), accession 0001104659-26-075184
• 8-K, 2026-06-08 (annual meeting results), accession 0001104659-26-071384
• 8-K, 2026-03-24 (Wells Fargo fourth amendment), accession 0001104659-26-033893
• DEF 14A, 2026-04-24, accession 0001104659-26-048201
• Forms 4 (24 filings, 2026 YTD) and Forms 144 (21 filings since 2026-05-01), parsed from XML
• SEC XBRL companyfacts CIK 0000903651, retrieved 2026-07-29
• Alpaca Markets SIP daily bars and options snapshots, retrieved 2026-07-29