Palantir (PLTR): investment dossier

Working file as of 18 August 2026. Prices, multiples and scenario outputs are as of that date and have not been updated.

At a glance

---

Part 1 · Business model and the FDE mechanism

1. The core engine: the flywheel of land (win new customers) and expand (grow existing customers)

• Land phase: build an Ontology for a new customer. That means integrating the customer's data, scattered across its various systems (ERP, CRM, supply chain and so on), and mapping it into one unified business-object model. This is laying the foundation, the phase of intensive FDE (Forward Deployed Engineer) investment. It is costly and takes a long time.
• Expand phase: once the foundation is built, the building goes up on the same Ontology, with new modules and new workflows. This step is cheap and high in gross margin. It is the main source of profit, and it shows up as extremely high net dollar retention (NDR, 157% in Q2 2026).
• Flywheel: after the land inflection point, FDEs pull back from existing customers. Expand deepens there, while the FDEs go off to land the next new customer.

2. FDE is the moat, and also a hard limit on growth

• The real barrier around the Ontology is the dirty work at the business semantic mapping layer: translating how a particular company defines 'an order', 'a region' or 'a channel conflict' into a data model. This work is highly customised. The customer's own technical team cannot do it, so FDEs have to build it together with the business side.
• Key conclusion: Palantir has not shaken off the FDE constraint on land. Productisation tools such as Bootcamp and AIP lower the threshold for expand, not the threshold for land (building the foundation).

3. Industry templates can speed up land, but only the foundation layer

• Palantir really is building industry templates, or accelerators, through a mechanism called Ontology Compounding. The custom pipelines that FDEs build on site at customers flow back to headquarters and are abstracted into standard templates reusable within the same industry. So the more customers it does in one industry, the faster it lands the next (for example, from Coca-Cola to Pepsi).
• But the templates speed up only the standardisable layer of technical pipelines and data structures. The business semantic mapping layer is where the real talent bottleneck lies, and the templates cannot help with it. FDEs still have to build it with each customer, one by one. Every time Palantir enters a brand-new industry, it has to accumulate templates from scratch again.

4. AI is loosening this constraint, but has not yet broken through it

• AI is taking over much of the data-cleaning and pipeline-integration layer, the most labour-intensive part of building an Ontology. That amplifies the capacity of each FDE.
• CEO Alex Karp has said publicly (June 2026, cross-checked against multiple sources):
  • 'The goal is to get 10x revenue and have 3,600 people. We have now 4,100.'
  • Plans to 'largely freeze hiring', relying on AI to amplify capacity rather than on layoffs.
• Implication: if 'fewer staff, more revenue' is delivered, it would break through the revenue ceiling (AI amplifying land capacity) and the margin ceiling (falling absolute headcount means lower costs) at the same time, which would benefit both growth and profitability.
• Important caveat: Karp's goal of 10x revenue with headcount cut to 3,600 is a vision or guidance, not a delivered fact. There is a motive to talk up the valuation, so it must be checked against actual data.

5. The long-term milestone that would truly rewrite the valuation logic

• The question is whether AIP can automate the building of the Ontology. At present, AI FDE can do the data pipeline layer but not yet the layer of business semantic mapping and co-building with the customer. Once that layer is automated, the hard talent ceiling on land disappears and the whole growth logic will be rewritten. This has not happened yet.

6. The three metrics to watch (easier to measure, and quicker to warn, than asking how much TAM is left)

1. Net dollar retention (NDR): the gauge of expand headroom. The baseline is 157%. The real alarm is a fall below 120%, or a one-way decline over two or three consecutive quarters. A fall below 150% is not an alarm, because a high base entering the denominator will pull the figure back mathematically.
2. Net new customers: the gauge of land speed. US commercial customers in Q2 2026 were 653, up 35% year on year but only 6% quarter on quarter. The slowdown quarter on quarter is worth watching.
3. The scissors gap between revenue growth and headcount growth: the litmus test of whether FDE productisation has succeeded, and whether growth is ultimately product-driven or labour-driven. The wider the gap, the more revenue each FDE supports and the more the flywheel is spinning on its own. If the two move in lockstep, the company is still quietly piling on people. A sharper version is to check directly whether absolute headcount is really going down (towards Karp's 3,600 target), and whether it is in fact still rising despite the talk of cuts.

---

Part 2 · Valuation model

Rebuilt spreadsheet, 17 Aug 2026. It replaces an earlier version whose rows stepped 1, 3, 6, 10, 15, 21 (triangular exponents) and carried no year labels. That unlabelled grid is the origin of the three-versus-six-year error below. The rebuild uses explicit consecutive year rows (PE year 1 to PE year 10) and adds an R=10% column, so the trap is gone.

Forward P/E decay grid: 153 / (1+R)ⁿ, TTM P/E (year 0) = 153

Every cell shown reproduces the formula exactly.

The row highlighted in red on the sheet is year 3, and this time it really is year 3.

Three-year scenario blocks (exit multiples scale with growth, current market cap $430 billion)

All three blocks reproduce exactly (753 / 518 / 379; 20.5% / 6.4% / −4.1%). The sheet labels these exits 'PEG = 1'; on the three growth rates they work out at a PEG of 0.67 to 0.8, so if anything they are conservative.

⚠️ Horizon: the sheet settles it

The rebuilt sheet resolves an open question from an earlier estimate, and it resolves it against that estimate's headline number.

The earlier estimate claimed about 4.9x and $2.1 trillion over three years. The sheet's own three-year row says the base case returns 1.2x ($518 billion, +6.4% a year) and the bear case loses money (0.9x, −4.1% a year). Those are the honest three-year numbers.

Where 4.9x actually lives: solving 153 / 1.6ⁿ = 9.12 gives n = 6.0 years. The 3.0x / 4.9x / 9.4x ladder ($1.28 trillion / $2.12 trillion / $4.04 trillion) is a six-year result, applying the same 40/45/50x exits to the year-6 row:

Reaching 4.9x in three years would require about 155% annual growth (at a 45x exit) or an exit multiple of about 183x (at 60% growth). Neither is plausible, so the figure cannot be relabelled as a three-year one. It has to be six years.

Net: 'It is a six-year idea that looks like a mistake for the first three.' Our video script already stakes out the same position. Its three-year reality-check rows (the base case up about 20% in total, the bear case losing money) match the rebuilt sheet exactly, so script and spreadsheet agree.

---

Part 3 · The bull case, stress-tested (18 August 2026)

Source: a structured bull-versus-bear stress test (18 Aug 2026) that tested every link. Purpose: not to argue that PLTR will definitely go up, but to take apart every link in the bull case and, in the end, see clearly what exactly we are betting on, and what signals would tell us the bet is wrong. Core conclusion: external threats (competition, AI disruption, cheap substitutes) were essentially argued down to secondary risks on a five-year horizon. All of the real uncertainty converges on three endogenous questions, which only time and earnings data can answer.

3.1 Valuation method: from the flaws of PEG to the forward P/E trajectory

PEG is a crude metric. At the same PEG of 1, 50% growth at 50x P/E and 100% growth at 100x P/E are not the same. Growth compounds (it is exponential), while P/E is just a linear purchase cost. PEG measures something exponential with a linear ratio, so it inherently undervalues high growth.

Numerical check (starting EPS of 1, exit after 4 years at 20x P/E):

• A: 50% growth, bought at 50x. EPS reaches 5.1, market cap 101, annualised return about 19%
• B: 100% growth, bought at 100x. EPS reaches 16, market cap 320, annualised return about 34%

PEG's three blind spots:

1. Sustainability (duration risk): the higher the growth rate, the harder it is to sustain
2. The lifespan of growth: 20% running for 15 years may be worth more than 100% running for 2 years
3. The quality of growth: organic growth driven by ROIC or ROCE versus cash-burning, dilutive share issuance

Upgraded method: a multi-year forward P/E trajectory plus an exit anchor

• Hold the share price fixed and let only EPS roll forward at the respective growth rates
• Use each year's EPS to back out that year's forward P/E
• Judge that year's P/E against the P/E it should command at that time (the exit anchor)
• Three gaps must be filled. First, the exit anchor is the real point of judgement, and it depends on the remaining growth rate. Second, growth-rate decay must be discounted by hand. Third, build it as a trajectory to see the slope of the P/E collapse

3.2 Working through the PLTR numbers

Base data (checked first-hand):

• Current price about $174 (about 16% below the 52-week high of $207.52)
• Q2 2026 single-quarter GAAP net income of $1.07 billion, diluted EPS of $0.41 (last year $0.13, up more than 3x)
• Q2 2026 revenue of $1.94 billion, up 93% year on year, with US commercial up 149% and gross margin of about 85%
• NDR of 157% (up a further 700 basis points quarter on quarter from Q1's 150%, a record high)
• FY2026 revenue guidance raised to growth of 82% year on year
• TTM GAAP diluted EPS of about $1.20 and current TTM P/E of about 145x
• Diluted share count of about 2.6 billion shares

Growth-rate scenario (the bull-case assumption, not a baseline): Y1 +100% / Y2 +100% / Y3 +75% / Y4 +60% / Y5 +45%

Year-four exit anchor (at PEG of 1):

• Basis one (Y4 forward growth rate 45%): fair P/E of 45, fair share price of $604.8, 3.48x over 4 years, annualised 36.5%
• Basis two (Y4 same-year growth rate 60%): fair P/E of 60, fair share price of $806.4, 4.63x over 4 years, annualised 46.7%

Even if Y4's $604.8 is discounted back to today at a high 25% discount rate, it is still worth about $248, more than 40% above the current price of $174.

3.3 Absolute profit and implied revenue: a reality check

Y5 absolute net income: $50-58 billion (including 3% annual dilution), roughly the same tier as Meta today (about $62 billion). It requires PLTR to reach Meta-level absolute profitability five years from now.

Backing out revenue (assuming a 58% net margin):

Key insight: TTM revenue of about $10 billion rising to $87.4 billion in Y5 is 8.7x, or annualised revenue growth of about 54% over five years. Compared with the current actual growth rate of 93%, this is a path of clear deceleration. Mathematically, it is not aggressive.

The bet boils down to something very pure:

• The revenue side (decelerating to 54% annualised) is the most credible
• The net margin side (30% to 58%) is the core bet

How realistic a 58% net margin is:

• It exceeds the historical peaks of Microsoft (about 35%) and Meta (about 35%), and sits at the upper end compared with Visa and Mastercard (50-55%)
• But it does not require the ceiling of software history (only 65% would require collecting extreme monopoly rents)
• Logic chain: a high net margin comes from extreme operating leverage (fewer staff, more revenue), which comes from AI amplifying FDE capacity
• Echoes Karp: 10x revenue with 3,600 employees (currently 4,100), plus the plan to 'largely freeze hiring' and rely on AI to amplify capacity
• Caveat: Karp's goal of fewer staff and more revenue is a vision or guidance, not a delivered fact, and there is a motive to talk up the valuation

3.4 The moat debate: taking the threats apart one by one

3.4.1 Core asset: the Ontology layer

The digital twin plus AI brain framework: the Ontology is the enterprise's digital twin. It models all of the enterprise's data, business processes and decision logic, and it takes root in the enterprise's central nervous system. Before, it was only a body. Only with AI added does it have a brain. The problem for competitors (Microsoft, Google, Databricks) is that they don't even have the body. They provide tools, models and data warehouses, but not the layer that semantically models the enterprise's business logic. Connect a large model to raw data without an Ontology and it can only chat. It cannot deliver decisions.

The Ontology is a time barrier built up over 20 years. NDR of 157% is ironclad proof that customer dependence is deepening.

3.4.2 Switching costs higher than Visa/Mastercard (the bull case)

Bull case: PLTR will become a monopoly (whereas Visa and Mastercard, V/MA, are a duopoly), with greater customer dependence. Without the Palantir system, enterprise teams simply cannot function.

Points of agreement: switching costs here crush V/MA's by an order of magnitude. Visa's moat is network effects (limited switching pain per customer). PLTR's is deeply embedded lock-in: ripping it out would force the enterprise to redo its digital transformation. This is a mission-critical system.

Two reality checks:

1. V/MA's 50-55% net margins come not from monopoly pricing power, but from zero variable cost. Visa's take rate is actually very low. The high net margin comes from the marginal cost of incremental transactions approaching zero. PLTR's incremental revenue still requires incremental FDEs to land new customers and build Ontologies, which is labour-intensive. A monopoly position does not mean a high net margin. Cost structure is what determines net margin. Whether PLTR reaches Visa-level net margins depends on whether AI can turn building an Ontology from labour-intensive work into something with near-zero variable cost.
2. 'Monopoly' is future tense. It does not hold today. Today PLTR faces attack from above (Microsoft Fabric/Copilot, Google, Databricks, Snowflake), substitution from below (enterprises building their own) and competition from peers (C3.ai). It is a leader in a niche, not a monopoly of the whole market.

3.4.3 FDE constrains only land, not operations (the bull's point, accepted)

The bull's rebuttal: FDEs are needed in large numbers only during the early land phase. Once the system matures, the enterprise's own technical staff maintain it. AI will greatly improve FDE efficiency.

Correction accepted:

• FDE is a one-time build cost (building the house), not an ongoing operating cost (property management).
• Expand at mature customers relies on the productised tools AIP and Bootcamp. It is driven by the enterprise itself and does not consume FDEs. Most of the increment in the 157% NDR is productised expand, and this part is rent collection with near-zero variable cost.
• AI really is raising the land capacity of each individual FDE.

Residual constraint: FDE is still the constraint on customer-acquisition speed (land). Growing 8.7x in five years depends heavily on land speed, and AI cannot yet replace the hardest layer of building an Ontology (translating the customer's messy business logic into a data model). The constraint has not disappeared. It has moved, from a service constraint to a customer-acquisition speed constraint.

3.4.4 The competitor debate

Bull case:

• Google, Microsoft and Databricks have no Ontology layer and cannot build one in three to five years
• OpenAI and Anthropic execute well, but they are fighting on too many fronts and their energy is spread thin
• Enterprises naturally resist being locked in by model companies (by analogy, using OpenClaw rather than Claude Cowork), so PLTR's model neutrality (any LLM can be plugged in) is actually a selling point

Points of agreement:

• Competitors really do lack the body layer, which is an insightful judgement
• Enterprises resist being locked in by model companies: this is sharp and correct, and it is a structural obstacle for OpenAI and Anthropic in building the enterprise middle layer

Two flaws:

1. 'No competitors' overstates it: the threat comes not from copying the Ontology but from going around it. Microsoft Fabric plus Copilot covers 70% of use cases with a lighter, cheaper solution, and PLTR's TAM gets nibbled away from below. Monopolising the high end is not the same as monopolising the market.
2. Underestimating the value of breadth: the giants are pushing forward on the momentum of the entire AI wave and may build the enterprise layer along the way, and PLTR cannot keep up with the pace of model iteration (its brain is rented)

3.4.5 The land-grab thesis: the most essential point of the whole bull case

Bull case: five years from now, most valuable enterprises will already have grown onto PLTR, all connected through it as their decision hub and unable to do without it.

Agreed: a classic case of first-mover lock-in and winner takes all. Each enterprise installs only one enterprise-grade central nervous system (an exclusive asset). Whoever first builds the Ontology into an enterprise and grows it into the decision hub locks in that customer for an extremely long time. It is a window-of-opportunity business. Whoever grabs the most high-value large enterprises in the next three to five years wins the right to collect rent for the next twenty years. Every customer landed is a plot of land permanently taken out of rivals' potential territory. Once the land has been grabbed, growth shifts from labour-intensive land to zero-marginal-cost expand, which is rent collection, and net margin rises naturally. The whole thesis closes its own loop here.

The only flaw: is the land big enough? Today there are only 653 US commercial customers. If only two to three thousand are grabbed in five years, what gets locked in is land that is high-value but not big enough. That is enough to support a great company, but not necessarily enough to support an 8.7x valuation in five years and Meta-level profits. The bet is not on whether PLTR will win, but on whether the land it grabs is big enough to support today's price.

3.4.6 Cheap solutions will not cause data fragmentation (the bull case refutes the diversion threat)

Bear-case scenario one: large enterprises' peripheral use cases (60%) are diverted to Fabric plus Copilot, compressing the share of wallet per customer.

The bull's three-part rebuttal (escalating):

1. Migration has a cost: connecting the data, maintaining a second system and bearing the risk of fragmentation. As long as PLTR's premium is below the migration friction, enterprises won't bother to move.
2. Charging for outcomes: PLTR's pricing is 'the enterprise earns 100 more, PLTR takes 30, the enterprise keeps 70'. Payment is tied to benefit, so enterprises have no incentive to flee. TAM expands from IT budgets to corporate profits.
3. Data sovereignty (the decisive blow): if the marketing department uses a cheap system, then once the data is fragmented, the CEO can no longer ask cross-departmental, company-wide questions in PLTR. The cost of fragmentation is borne not by the department that decided to save money but by the top decision-makers, and the decision-makers have the power to block it.

Refutation accepted: scenario one is downgraded from a competitive threat to something that essentially does not hold. It forms a centripetal pull on data, a network effect: each new piece of data that connects makes the next one more inclined to connect. Lock-in does not mean that switching is unaffordable. It means that no one will even propose switching, because proposing it means creating a blind spot for the CEO.

The only residual flag: the premise carries one condition, 'as long as PLTR's pricing isn't outrageous'. The ultimate Achilles' heel is not competitors or AI but PLTR's own pricing discipline (management behaviour risk).

3.4.7 Clarifying the pricing structure: a hybrid of outcome and usage

The bull's observation (accurate): as things stand, pricing is a hybrid of outcome-based and usage-based.

The three layers of the hybrid:

1. Platform subscription base (usage- or seat-based): predictable recurring revenue
2. Bootcamp and deployment services (project-based): fades out as maturity increases
3. Value-linked portion (outcome-based): an embryonic form of 'earn 100 more, take 30', but not pure profit-sharing, and closer to value-anchored, tiered usage expansion

Why a hybrid rather than pure outcome-based pricing: the fatal weakness of pure outcome-based pricing is attribution. The rational implementation is for the base subscription to lock in cash flow with certainty, and for the increment to share the upside through value-anchored usage expansion. In essence, it delivers the effect of outcome-based pricing in the form of usage-based pricing.

Calibration:

• The upside: steadier and more sustainable than pure outcome-based pricing. This is exactly the mechanism behind the 157% NDR: not price increases, but the deeper the use, the greater the usage.
• Needs discounting: the strong claim of 'earn 100 more, take 30' should be reined in. In reality, PLTR captures part of the value indirectly through usage growth, and does not directly take 30%. The pricing ceiling is anchored to IT and data budgets plus a value premium, not to the pool of all corporate profits.
• NDR is the single best reading of whether this hybrid mechanism is healthy.

3.5 The ultimate AI/ASI threat: the intelligence-versus-tools framework

Bear-case scenario two: model commoditisation will kill PLTR. AI becomes strong enough to get by even without an Ontology, and the value of the body is diluted.

The bull's strongest argument: no matter how high intelligence gets, it cannot replace tools. What tools do is lower the cost and difficulty of intelligence solving problems, and they are not subsumed by intelligence.

• A person with an IQ of 200 working through a complex ledger can do it without Excel, but the time it takes explodes (a cost problem)
• Deny them even scratch paper, and anything slightly complex simply cannot be solved (not enough working memory, which makes it a capability boundary)
• Tools are intelligence's external memory and compute offloading, and they are part of capability

Applied to PLTR: the Ontology is that scratch paper and that Excel: the external working memory of enterprise-grade AI. Without an Ontology, even the strongest AI (even ASI), facing a large enterprise's dozens of systems and petabytes of messy data, is like the IQ-200 person denied scratch paper. The stronger the AI, the more it needs good tools to unleash its capability. A high-quality body is an efficiency multiplier for intelligence, not a substitute. Scenario two is downgraded from fatal to secondary.

But there is one crack: Excel is a general-purpose tool (it belongs to no one), whereas the Ontology is proprietary to PLTR. The framework shows that the AI era will certainly need a tool layer like the Ontology, but not that this tool layer must be PLTR's. If ASI is strong enough, it may quickly generate a good-enough Ontology on its own. One cannot assume both that ASI is strong enough to need tools and that it is too weak to build them.

The bull's key qualifier: human-in-the-loop

The bull's rebuttal: the possibility of ASI building an Ontology itself exists, but building an Ontology embeds a large number of decisions that need a human to sign off (definitions of business concepts, whether fields are the same, exception rules, data trustworthiness). These bottlenecks lie not in AI intelligence but in human judgement and consensus. As long as enterprises still insist on human-in-the-loop, the speed at which ASI builds an Ontology is capped by the speed of human sign-off. The friction shifts from whether the AI can work it out to whether humans are willing, and able, to reach consensus. ASI cannot eliminate that layer of friction, because it is an organisational and accountability problem, not an intelligence problem.

Accepted: the timeline changes from when ASI will be smart enough (possibly soon) to when enterprises will dare to let AI make business-definition decisions with full authority (much slower, because it involves trust, accountability, regulation and law). The probability of enterprises removing humans from the decision loop at scale within five years is extremely low, so the five-year investment thesis is safe.

The final twist (where the analysis ends up): the moat holds on the premise that enterprises insist on human-in-the-loop. But the entire direction of AI development is to keep squeezing humans out of the decision loop. If competitive pressure forces enterprises to remove human-in-the-loop gradually, then the premise that human friction protects PLTR disappears, and the floodgates open for ASI to build the Ontology on its own. This would be a prisoner's dilemma: if the first enterprise that dares to let AI make autonomous decisions gains an overwhelming advantage in speed and cost, the others are forced to follow.

PLTR's moat rests, in essence, on the fact that humans are indispensable in enterprise decision-making. It is safe as long as humans remain in the decision loop. Its ultimate risk is not any competitor, but the AI wave itself one day squeezing humans out of the decision loop. That day will also be the day the Ontology's human friction disappears and the moat runs dry. That day will not come within five years, but it means PLTR should not be a permanent heirloom, but a long-term holding that requires continuous monitoring of whether human-in-the-loop is loosening.

3.6 Valuation positioning and the final convergence

3.6.1 The current price: not cheap, but not expensive

The current price of $174 roughly corresponds to this: the market has already priced in the high-speed but normally decaying scenario, but has not yet priced in the aggressive curve.

• It is not bargain hunting (a bargain is when the market hasn't seen the growth, whereas the whole world is watching PLTR's growth)
• Nor is it chasing a bubble at the highs (a bubble is when the price has already run ahead of the most optimistic assumptions, whereas the current price has not yet reached the aggressive curve's fair value in the $600 range)
• It is paying a fair price for a good outcome that is very likely to materialise, while getting a free upside option on delivery beyond expectations

Trading discipline: this is not suited to going all-in with a heavy position, because there is no margin-of-safety discount. It suits holding according to plan, or buying in tranches, because the odds are asymmetric, with upside far greater than downside.

3.6.2 Three failure scenarios (all taken apart or downgraded)

1. Diversion to cheap solutions: refuted by migration costs, outcome-based pricing and data sovereignty, and downgraded to something that essentially does not hold. Residual: pricing discipline
2. Commoditisation of model capability (ASI): downgraded to secondary by the intelligence-versus-tools framework. Residual: whether human-in-the-loop loosens
3. Overstretched valuation: not cheap, but not expensive, and dependent on the speed and sustainability of growth. Current judgement: reasonable

3.6.3 All uncertainty ultimately converges on three endogenous questions

1. How big is the land? This is the TAM boundary. Watch: net new customers and contract value per customer (TCV/customer)
2. How fast can it be grabbed? This is the pace of delivery against an overstretched valuation. Watch: the scissors gap between revenue growth and headcount growth
3. Will it self-destruct? This covers pricing discipline, SBC convergence and whether fewer staff and more revenue is actually delivered. Watch: whether NDR has peaked, SBC as a percentage of revenue, and whether absolute headcount is moving towards the 3,600 target

3.6.4 Quarterly earnings monitoring dashboard

• NDR: the single metric to watch most closely. Q2 2026 was 157%. The real alarm is a fall below 120%, or a one-way decline over two or three consecutive quarters (not a fall below 150%). Note that NDR is measured on a trailing 12-month basis. Next year the high base enters the denominator and it will fall back mathematically, so don't extrapolate 157% as a steady state.
• Pace of net new customers: tests whether land is accelerating (progress in the land grab).
• The scissors gap between revenue growth and headcount growth: the litmus test of product-driven versus labour-driven growth. If the gap opens up, PLTR is becoming a zero-marginal-cost machine. If it fails to open up, PLTR is still a high-end project outsourcing company.
• Absolute headcount: whether it is really going down (the direct check on whether Karp's goal of fewer staff and more revenue is being delivered).
• Trend in SBC as a percentage of revenue: whether it falls quarter by quarter from about 14%. This is the direct leading indicator of whether net margin can reach 58%. Note that high share prices in a bull market push up the cost of newly granted SBC, so convergence may be slower.

3.6.5 The SBC convergence debate

Bull case: once the company has matured, the moat is established and customers are locked in, it won't need so much equity compensation, and paying cash will do.

Calibration (the direction is right, but don't treat it as certain):

• SBC as a percentage of revenue will very likely go down, so the direction is right.
• But two preconditions need watching. First, the absolute amount of SBC must not balloon out of control (PLTR has a prior record: historically SBC was 20-30% of revenue, the bears' main line of attack). Second, the share-price rally needs to level off for a while (SBC is expensed at fair value at the time of grant, so high share prices in a bull market actually push SBC up, contrary to the intuition that it should fall once the company matures).

3.7 Summary: what exactly are we betting on

The core bull-case logic is solid: the Ontology's 20-year time barrier, model neutrality, enterprises' resistance to being locked in by model companies, the centripetal pull of data sovereignty, and the fact that the stronger the intelligence, the more valuable the tools.

The bet boils down to something very pure:

• Revenue decelerating to 54% annualised (the most credible)
• Net margin from 30% to 58% (the core bet; it depends on whether AI can take the human labour out of building the Ontology)

The ultimate uncertainties, which cannot be eliminated by logic and can only be answered by time:

1. Medium term (within five years): how big the land is, how fast it can be grabbed, and whether PLTR will self-destruct through pricing greed or runaway SBC. These are monitored via the quarterly dashboard.
2. Long term (ten years or more): whether human-in-the-loop gets squeezed out of the decision loop by the AI wave. This is the ultimate Achilles' heel of PLTR's moat, which right now can be neither falsified nor verified, only continuously observed.

Investment profile: the current price is not cheap but not expensive, the odds are asymmetric to the upside, and it suits holding according to plan or buying in tranches rather than going all-in with a heavy position. PLTR is not a permanent heirloom, but a high-quality long-term holding that needs to be checked against the dashboard every quarter, with long-term monitoring for signals of its ultimate Achilles' heel. When the industry shows early signs of enterprises beginning to let AI make autonomous decisions at scale and gaining an overwhelming advantage, that is the time to re-examine the entire thesis.