Sandra Chen has spent eleven years on due diligence teams evaluating healthcare technology investments. She has a rule: if the pitch deck shows one cost-to-serve number, she multiplies it by two and works from there. Single-number economics in healthcare technology are almost always wrong. They average across populations that should not be averaged, they assume support costs that decline when they actually increase, and they treat infrastructure as a one-time expense when it is always ongoing.
When she opened the BlueMirror data room, she expected a single cost-to-serve figure and a single margin projection. She found six deployment paths, each with its own cost profile, and a cost floor that the company described as conservative rather than optimistic. She did not multiply by two. She started reading.
The six deployment paths and their cost profiles#
BlueMirror’s three-zone compute architecture, described in BMT-09.01, produces six distinct deployment configurations. Each subscriber is on one of these paths based on her hardware situation and the infrastructure available in her region. The unit economics vary by path because the cost components shift: a subscriber with a dedicated Local Pane device offloads inference from the cloud, reducing Zone 3 cost but adding device amortization. A subscriber without a device pushes more workload to Zone 3, eliminating hardware cost but increasing cloud inference spend.
The six paths and their per-subscriber monthly cost at scale:
| Path | Configuration | Device Amort. | Zone 2 Share | Zone 3 Inference | Software/Ops | Support | App/Bandwidth | Total Range |
|---|---|---|---|---|---|---|---|---|
| A | Z1-Dedicated + Z2 + Z3 | $5–7 | $5–7 | $2–5 | $3–5 | $5–12 | — | $20–36 |
| B | Z1-Dedicated + Z3 | $5–7 | — | $8–14 | $3–5 | $5–12 | — | $21–38 |
| C | Z1-Phone + Z2 + Z3 | — | $5–7 | $2–5 | $3–5 | $5–12 | $1–2 | $16–31 |
| D | Z1-Phone + Z3 | — | — | $8–14 | $3–5 | $5–12 | $1–2 | $17–33 |
| E | No Z1 + Z2 + Z3 | — | $6–8 | $3–6 | $3–5 | $5–12 | — | $17–31 |
| F | No Z1 + Z3 | — | — | $10–18 | $3–5 | $6–14 | — | $19–37 |
Path A is the full-stack deployment: a dedicated Local Pane device in the subscriber’s home, a shared Zone 2 regional node at a PACE facility or care agency, and Zone 3 cloud for complex reasoning. Path F is the other end: no dedicated device, no regional node access, Zone 3 carries the full inference workload. The remaining four paths fall between them.
Each path reflects a real population segment. Path A concentrates among PACE enrollees and Medicare Advantage subscribers where the institutional payer funds the Local Pane device and the PACE facility or care agency hosts the Zone 2 node. Path C concentrates among younger aging adults (60–70) who already own capable smartphones and live in regions where a Zone 2 node is deployed. Path F concentrates among subscribers in rural areas without Zone 2 coverage, or among the oldest cohort who do not use personal devices and whose interaction with the system happens through IVR, caregiver-mediated interfaces, or basic client modes. Paths B and D emerge where a subscriber has a device (dedicated or phone) but no regional node has been deployed in her area. Path E serves subscribers who access a Zone 2 node through a care agency or community center but do not have a personal device at home.
The ranges within each path reflect real variation. Support costs for a 62-year-old in her first year differ from support costs for an 81-year-old in her seventh year whose cognitive needs have increased. Zone 3 inference costs for a subscriber who asks twelve questions a day differ from a subscriber who interacts three times. The ranges capture what the system actually encounters across the subscriber population, not what an average scenario predicts.
The support line deserves attention because it is the cost component that does not decline with scale or maturity. Device amortization is fixed. Inference cost declines as SLMs train and caching improves. Software and operations cost spreads across subscribers. But support for a population that is aging progressively (some of whom will experience cognitive decline, mobility loss, and increasing complexity of needs) trends upward over the subscriber’s tenure. The $5–12 range on most paths (and $6–14 on Path F, where the absence of a device increases the support burden) reflects this reality.
Why the cost variation is smaller than expected#
Sandra expected Path A to be significantly more expensive than Path F. Device amortization adds $5–7/month that Path F avoids entirely. But the cost difference narrows because the architecture trades one cost component for another.
Path A’s Local Pane device handles 15–20% of all inference locally, including the privacy-critical models that would otherwise run in Zone 3. Path A’s Zone 2 regional node handles another 55–60%. Zone 3 carries only 5–10% of queries for Path A subscribers, at $2–5/month. Path F pushes the full workload to Zone 3. No local device. No regional node. Zone 3 inference rises to $10–18/month because it carries every query that would have been handled locally or regionally.
The device amortization that makes Path A appear expensive is offset by reduced cloud cost. The absent hardware that makes Path F appear inexpensive is offset by higher cloud spend. Path B, which skips Zone 2 but retains the Local Pane, sees Zone 3 carry the Zone 2 workload at $8–14/month. Path E, which has Zone 2 access but no device, shifts more load to the regional node, increasing its per-subscriber share to $6–8.
The result is approximate path neutrality at scale. The architecture’s cost structure converges across paths because each architectural choice (cheaper hardware, regional pooling, cloud absorption) trades against the others. No single path breaks the model. No single path subsidizes the others at the unit level.
This convergence was not obvious during the original architecture design. The initial assumption was that Path A would be significantly more expensive and Path F significantly cheaper, which would create pressure to steer subscribers toward cheaper paths, a dynamic that would undermine the equity commitment. The three-zone revision revealed that the cost-shifting mechanism is approximately symmetric: what you save on hardware, you spend on cloud; what you save on cloud, you spend on hardware. The variance across paths at scale is approximately $8–12/month at the midpoint, small enough that path-uniform pricing is commercially sustainable rather than a cross-subsidy arrangement.
This convergence is what makes the equity commitment commercially viable rather than aspirational. BlueMirror can offer the same product at the same price across all paths because the cost to serve is approximately the same. The hardware substrate differs. The cost does not diverge enough to require path-based pricing.
The consumer rate schedule#
The subscriber pays the same rate regardless of her deployment path.
| Period | Monthly Rate | Rationale |
|---|---|---|
| Year 1–3 | $100 | Cold-start overhead: high support needs, device provisioning for Path A/B subscribers, Zone 2 infrastructure buildout, SLMs not yet trained on individual context |
| Year 3–5 | $70 | SLMs trained, context rich, support stabilized, infrastructure amortized |
| Year 5+ | $50 | Convergence with sustainable list price |
| Over-70 loyalty (3+ continuous years) | $35 | Cost floor, no margin |
The over-70 loyalty rate is not a discount. It is the cost floor: the rate at which BlueMirror can serve a subscriber without losing money. A subscriber who joined at 67 and stayed continuously for three years reaches the loyalty rate at 70. She pays $35/month for as long as she remains a subscriber. BlueMirror earns no margin on her service. The rate honors a duration commitment by eliminating the margin, not by reducing the service.
The rate schedule is not a promotional ladder. Each step reflects a genuine change in cost-to-serve. In years one through three, the system is learning the subscriber: the SLMs are training on her patterns, the preference model is calibrating, the support team is handling her onboarding questions, and for Path A and B subscribers the device has been provisioned and installed. These are real costs that front-load the relationship. By year three, the SLMs are trained, the context is deep, the subscriber rarely needs human support for routine interactions, and the device (if applicable) is operational and amortizing. The $70 rate reflects this reduced cost. By year five, the marginal cost of serving her has declined further. The $50 rate is the sustainable list price at which BlueMirror operates profitably on every path at scale.
The path-level cost variation is absorbed by BlueMirror, not passed to the subscriber. A Path A subscriber at $100/month in year one generates more margin than a Path F subscriber at $100/month in year one, because Path A’s cost-to-serve midpoint is slightly lower at scale. But the subscriber never sees this. The pricing is path-independent because the service is path-independent.
Lifetime value calculations across paths#
Four scenarios model the range.
Scenario 1: Self-paying subscriber, age 60, Path A, 10 years. Revenue: $100 × 36 months + $70 × 24 months + $50 × 48 months = $7,680. Average cost-to-serve across the lifetime as SLMs mature and marginal inference cost declines: approximately $48/month in the early years, falling as the system learns her. Average gross margin: approximately $30/month across the full tenure. This is the high-margin subscriber.
Scenario 2: Self-paying subscriber, age 67, Path D, 15 years. She uses her smartphone as the Zone 1 device. No Zone 2 node in her region. Revenue: $100 × 36 months + $35 × 144 months = $8,640. She transitions to the over-70 loyalty rate at year three. Average cost: approximately $52/month (higher Zone 3 inference, increasing support needs as she ages). Average margin: approximately $25/month in the early years, near zero after the loyalty rate takes effect. The first three years fund the margin. The remaining years are at-cost.
Scenario 3: Institutionally funded subscriber, age 74, Path A, 10 years. A Medicare Advantage plan pays the institutional rate of $50/month. Revenue: $6,000. Cost-to-serve: approximately $36/month average. Margin: approximately $14/month. Steady, predictable, no consumer acquisition cost, no churn risk as long as the MA plan maintains coverage.
Scenario 4: Viability Gap Fund subscriber, age 72, Path F, 8 years. The Gap Fund covers $25/month, she pays $10/month. Total received: $35/month × 96 months = $3,360. Cost-to-serve: $35/month. Margin: zero. By design. The clinical outcomes data generated by this subscriber has value for institutional fundraising, regulatory positioning, and equity reporting, but the unit itself does not generate profit.
These four scenarios are not edge cases. They represent the four major subscriber archetypes that compose the blended business: the self-paying early adopter (Scenario 1), the self-paying subscriber who ages into the loyalty rate (Scenario 2), the institutionally funded subscriber who arrives through a channel partner (Scenario 3), and the Gap Fund subscriber who would otherwise be unserved (Scenario 4).
The blended business across these scenarios is profitable because the channel mix is weighted toward higher-margin segments. At scale, approximately 85% of subscribers arrive through institutional channels (Scenario 3 analogues), which provide steady revenue at moderate margin. Approximately 5% are direct-to-consumer (Scenarios 1 and 2), which provide higher margin early but transition to lower margin or cost-floor rates over time. The remaining 10% are provider-mediated or Gap Fund supported. The weighted average margin across the full subscriber base at scale is approximately 40% gross, driven primarily by institutional channel volume.
The model does not rely on cross-subsidy from one path to another at the unit level. It relies on funding-source diversification at the channel level, which BMT-10.02 describes in detail.
Why duration beats extraction#
The five-year subscriber at $50/month is more profitable to serve than the year-one subscriber at $100/month. The reason is cost decline.
By year three, the P-RLHF preference model (BMT-02.05) has learned the subscriber’s communication preferences, response style, and interaction patterns. The system generates fewer irrelevant responses, fewer clarification loops, fewer escalations to human support. The SLMs that serve her domain needs have been fine-tuned on her context: her medication list, her financial patterns, her home layout, her social connections. The marginal cost of the next query is lower because the system has already learned what she needs and how to deliver it.
Acquisition cost for a five-year subscriber is zero. She is already here. Support cost has stabilized. Context is deep. The descending price schedule retains subscribers who would churn at a flat $100/month rate. The price decline is not a concession; it is a recognition that the relationship has matured and the cost structure has changed. She is paying less because she costs less to serve, and because her continued presence generates compounding returns: her outcomes data strengthens institutional fundraising, her P-RLHF model improves the SLM training corpus, and her satisfaction reduces the acquisition cost of future subscribers through referral.
A competing platform that offers her a lower price cannot replicate three years of P-RLHF learning. The switching cost is genuine accumulated value: the system knows her. It knows that she prefers direct answers over explanations, that she takes her metformin at 7:15 AM with breakfast, that her daughter calls on Tuesdays and her son calls on Sundays, that she is anxious about her property taxes every October. That knowledge is non-transferable. It is not an artificial lock-in mechanism. It is the product working as designed.
This duration advantage holds across all deployment paths. The cost-decline mechanism is SLM training and context depth, both of which apply to every subscriber regardless of which zones she has. A Path F subscriber’s models train on her interactions just as a Path A subscriber’s models train on hers. The inference runs in different zones. The learning compounds identically.
The cost floor is real, not aspirational#
The $35/month cost floor is the rate at which BlueMirror can serve a subscriber on any path without losing money. It accounts for device replacement cycles (for paths that include dedicated hardware), ongoing model updates distributed across the three zones, increasing support needs as cognition changes over time, compliance and security operations, and Zone 2 infrastructure maintenance where applicable.
The cost floor was originally projected at $20/month. That projection assumed in-home GB10 deployment (pushing hardware cost to a one-time purchase rather than ongoing amortization), and it assumed support costs would decline as the system learned the subscriber. The first assumption was wrong because the revised three-zone architecture distributes compute across shared infrastructure with ongoing operational cost. The second assumption was wrong because support needs for the over-70 population with progressive cognitive change increase, not decrease. Customer support is the cost line that moves against the subscriber’s tenure, not with it.
The corrected cost floor of $35/month reflects both revisions. It is conservative in the sense that the midpoint across all paths is closer to $25–28/month, and a Path C subscriber in a mature market with a well-utilized Zone 2 node may cost as little as $16/month to serve. But the floor must cover the subscribers who cost more: the Path F subscriber whose Zone 3 inference load is heavy, the 85-year-old whose support needs require human intervention weekly, the rural subscriber whose Zone 2 node serves only twenty people instead of seventy-five. The $35/month figure covers all of them. That is what makes it a floor.
Cross-References#
The Three-Zone Architecture (BMT-09.01). The six deployment paths and the three-zone compute model that produces the cost variation analyzed here.
The Viability Gap Model (BMT-10.02). How the five funding layers close the gap between subscriber ability to pay and cost to serve.
The Retention Flywheel (BMT-10.05). The compounding dynamics that make duration more valuable than extraction.
What the System Learns (BMT-02.05). The P-RLHF mechanism that drives the cost-decline curve as subscriber tenure increases.