NEOMORPHIC™ SSI · GLOBAL LICENSING BUSINESS PLAN Trinity Sky · Palm Beach, Florida · 8 September 2026

Store more. Move less. Make AI memory more useful. License the improvement across devices and infrastructure.

The world is committing trillions of dollars to technology while data storage, transmission and memory constrain what that investment can deliver. Trinity Sky’s opportunity is to license the technology that improves those economics—through the companies that build chips, devices, cloud platforms and national digital infrastructure.

Phoenix targets the cost of storing and moving eligible data through a proprietary lossless encoding implementation. NEOMORPHIC Memory preserves, retrieves and verifies relevant evidence so AI can use knowledge efficiently as information changes. Together, the products address physical data cost and the repeated work—and accuracy loss—created by inadequate AI memory.

A global hardware and software licensing business

The named target accounts are MediaTek, Apple, NVIDIA, Google, AWS, IBM, Intel and Microsoft, alongside memory suppliers, design-IP partners and government programs. The proposed revenue model combines integration fees, annual minimums, shipped-component royalties and recurring active-device or infrastructure entitlements. These are prospective licensing relationships.

Preserve the commercial ambition and show the larger path

The management lead plan targets $155M revenue in Year 1 and $1.605B in Year 3. The new global-IP expansion case reaches approximately $5.22B in Year 3 and $47.34B in Year 10 through explicit adoption, unit-price and renewal assumptions. Both cases contain 120 monthly periods; their revenues are presented separately.

Initial seed sought: $20M. Customer receipts support the commercial operating program. The plan includes works cited for market scale, prospective customer platforms and industry licensing precedents, with product evidence and financial assumptions identified separately.

Companion presentation: 3sky.ai/deck

Investor confidential · Trinity Global Partners LLC (EGD33). Projections are conditional business-planning scenarios, not contracted revenue or investment returns.

---

01 / The investment case

NEOMORPHIC combines two opportunities: reducing the infrastructure cost of data and improving the usefulness of AI knowledge. The goal is a significant licensing position in markets where information is stored, transmitted, retrieved and used continuously.

The launch plan starts at commercial scale

The lead case targets $155M company revenue in Year 1, $611.7M in Year 2 and $1.605B in Year 3. Phoenix contributes $115.5M, $473.1M and $1.262B. These targets are supported by a disclosed enterprise, operator and OEM revenue build, with the sales, implementation and working capital needed to deliver it.

Distribution changes the growth equation

Enterprise licenses monetize high-value data estates. Operator contracts extend Phoenix across eligible network flows. Optional SDK and OEM agreements embed paid implementations into platforms and covered shipments. These channels begin in Year 1 in the management plan, enabling enterprise sales and broader distribution to grow together.

A larger IP licensing opportunity

The $143B management transmission-pool reference supports $1.43B annual vendor revenue at 1% capture. Hardware, deployed AI memory and sovereign entitlements extend beyond that single pool. The new global-IP case models these opportunities through named target cohorts and unit prices, reaching $47.34B of company revenue in Year 10.

Defensibility combines algorithmic performance, supported implementations, customer integrations and IP. Compression must improve the complete data path, and fast recall must return the correct evidence. Current-release validation strengthens the commercial case and supports proposed commercial deployment on the validated core.

---

02 / The economics of every data path

The opportunity extends across devices, memory systems, data centers and communications networks. Every repeated copy, transfer and retrieval consumes some combination of capacity, time and energy. An efficient implementation licensed into the platforms that perform those operations can earn revenue across successive product generations.

Size the paid product inside the spending

These figures describe overlapping markets at different supply-chain levels. They establish the scale of the problem. The licensing model then prices a defined product: a covered chip, an active device, an accelerator runtime, a controller, an enterprise deployment or a funded national program. This creates a route from broad industry demand to an auditable royalty statement.

Four pools of customer value

Storage: more useful data per physical byte and fewer replicated bytes. Transmission: fewer eligible bytes crossing a paid or constrained link. Inference and hardware memory: less eligible data movement and a smaller working set where the decoder can sustain the required rate. Persistent knowledge: more relevant evidence per request, less repeated context assembly and fewer stale or unsupported answers.

The same efficiency can support lower unit costs or more output from an existing budget. Procurement should measure both; capacity released does not automatically appear as a reduced invoice.

The strategic ambition is an IP position inside the infrastructure that stores, moves and uses the world’s information.

---

03 / The buyers are already investing at scale

The target accounts own the distribution channels and infrastructure budgets that can turn a successful implementation into a large recurring licensing business.

Amazon’s capital plan covers more than AWS, and NVIDIA’s revenue is supplier revenue. These measures are not summed as customer savings or a single addressable budget. Their relevance is the purchasing power and distribution reach of the proposed licensees.

Turn infrastructure scale into a purchasing decision

The first proposal to a hyperscaler should specify an eligible fleet, a measured reduction in cost per successful workload and a license whose price leaves the buyer a substantial share of that improvement. A $600 annual license across one million covered accelerators produces $600M recurring revenue. That is a proposed commercial illustration, not a published contract price.

At an illustrative $20,000 avoidable annual cost per accelerator, a verified 25% efficiency benefit creates $5,000 of annual capacity value. A $600 license leaves $4,400 per accelerator before any additional unmodeled integration expense. The inference example later in this plan shows the performance assumptions required to generate that benefit.

The scale comes from the platform agreement

One design win can reach multiple products, regions and refresh cycles. The agreement must identify deployment entitlements, royalty reporting, permitted subsidiaries and the next generation’s rights. Broader distribution then grows revenue without requiring a direct contract with every end user.

---

04 / Billion-dollar IP businesses already exist

Arm reported $4.92B of FY2026 revenue, including $2.61B of royalties and $2.31B of license revenue. That is an operating precedent for monetizing technology through other companies’ hardware distribution. It supports the business-model category, while NEOMORPHIC’s own design wins and prices remain to be secured. 50

Groq and NVIDIA announced a non-exclusive inference-technology licensing agreement in December 2025. The announcement confirms that a leading accelerator company can license external inference IP; it does not disclose a price to use as a NEOMORPHIC valuation or fee benchmark. 39

A repeatable commercial structure

Paid evaluation and integration. Deliver a workload benchmark, supported SDK and customer-specific integration under a fixed scope. Development fees fund engineering and establish the performance baseline.

Platform rights. License the optimized implementation, firmware or qualified hardware block for defined products, architectures and fields of use. Annual minimums support multiyear planning and reserve engineering capacity.

Running royalties. Charge per covered shipped component for embedded IP, or per active deployed unit per year for a supported runtime. Usage reports, audit rights and agreed exclusions make the entitlement measurable.

Generation expansion. Negotiate rights for additional chip families, operating systems, workloads and regions. A renewal or new generation can expand the paid scope without repeating the entire customer-acquisition process.

Preserve the IP while licensing its use

The proposed model licenses specified usage rights while retaining ownership of the underlying technology. Contracts separate background IP, customer-funded modifications, maintenance, source escrow and any exclusivity. Exclusivity should have a narrow field, a defined term and minimum commercial commitments.

An open format or reference implementation can support adoption. The paid value must reside in enforceable rights and differentiated implementation, performance, support or operational control. Patent filings, granted claims and freedom to operate require an IP register; this plan does not infer worldwide exclusivity from a portfolio description.

---

05 / The cost of the data bottleneck

Data creates expense at several points in its life: ingestion, storage, replication, backup, transmission, indexing, retrieval and repeated processing. These costs compound when multiple systems retain or move copies of the same information. Compression can reduce the byte component of that expense, but savings depend on where compression is applied and which costs are actually variable.

Storage: retain more useful information per unit of infrastructure

Phoenix targets datasets where exact reconstruction matters and enough redundancy remains to compress. Candidate entry workloads include structured telemetry, logs, recurring records, replication streams and machine-generated data. Each is a hypothesis to test, not a presumption that the same ratio applies across the portfolio.

Fortune Business Insights estimates the cloud storage market at $161.28B in 2025. Extending its published 2026 base and growth rate produces an illustrative $400.67B in 2030. This is a market for storage products and services, not a pool of immediately avoidable disk costs. 10

Transmission: reduce bytes before the expensive link

Compression can improve a constrained connection by sending fewer bytes, provided encoding and decoding do not consume more time than transmission saves. This is especially relevant to replication, remote facilities and eligible satellite or carrier workloads. The buyer cares about delivery cost and latency under its actual network conditions.

Ericsson forecasts 515 exabytes per month in 2031 including fixed wireless access, compared with 203 EB/month in 2025 on the same basis. Mobile traffic excluding fixed wireless access rises from 146 to 328 EB/month. These are different traffic scopes and must stay separate. 1

System efficiency: measure the whole path

The benefit calculation must include codec processing, memory consumption, metadata, retransmissions, integration and any existing minimum-spend commitments. The IEA projects data-center electricity demand reaching about 945 TWh in 2030. That scale makes efficiency commercially relevant; it does not establish an energy-saving percentage for Phoenix. 13

The product promise to validate: lower total cost per correctly stored or delivered unit of information.

---

06 / The memory and accuracy bottleneck

An AI system can retain an unchanged set of model weights and still become less useful as facts change, conversations grow or the information needed for a task falls outside the context it receives. Enterprises therefore need memory that persists, updates deliberately and supplies the right evidence at the time of use.

Bigger context is useful, but recall still needs testing

Research such as Lost in the Middle, RULER and NoLiMa shows that context length, information position and the type of retrieval task can affect performance. These studies motivate testing realistic memory workloads; their historical results are not a current leaderboard against NEOMORPHIC. 14, 15, 18

LongMemEval evaluates long-term memory through information extraction, reasoning across sessions, temporal reasoning, knowledge updates and abstention. Those dimensions provide a useful starting point for evaluating enterprise knowledge systems. 16

Distinguish three kinds of memory

Hardware memory is the physical RAM or accelerator memory available to an application. Inference working memory includes attention state and the key-value cache used while generating a response. Persistent knowledge memory stores information across sessions and model executions. NEOMORPHIC Memory primarily addresses the third category; Phoenix may assist the first two only where additional workload-specific tests establish useful compression.

PagedAttention addresses KV-cache allocation and fragmentation. FlashAttention reduces attention-related memory traffic. Quantization methods such as TurboQuant trade numerical representation against distortion. These approaches can complement persistent knowledge, and require different comparisons from a lossless general-purpose codec. 19, 20, 21

Accuracy is an operating requirement

The commercial pain is the cost of an answer that uses an outdated policy, omits a prior decision, confuses two entities or fabricates support. NEOMORPHIC’s proposed response is to preserve authoritative records, retrieve them efficiently, check their integrity and apply time and access rules before generation. The final answer still needs to be checked against that evidence.

The objective is dependable knowledge over time, measured by task performance and error rates. A precise memory architecture can reduce important causes of accuracy loss without claiming to eliminate every form of model or concept drift.

---

07 / Product architecture

NEOMORPHIC SSI is organized around an authoritative data layer, efficient representations and controlled access. This separates the record a customer relies on from the indexes and model outputs used to find and explain it.

flowchart TD
  A["Authoritative customer records"] --> B["Phoenix encoding"]
  A --> C["Memory indexing"]
  B --> D["Storage and transport"]
  D --> E["Exact reconstruction"]
  C --> F["Authorized, verified retrieval"]
  E --> G["Customer applications"]
  F --> G

Phoenix: the data-efficiency plane

The proposed codec encodes eligible data for storage or transport and reconstructs the original bytes at the destination. Its product boundary includes an encoder, decoder, container or framing format, version handling and telemetry. The decoder must remain available for the life of retained records, including customer export and disaster recovery.

Memory: the knowledge plane

Canonical records carry identity, version, provenance and integrity information. FHRR-derived representations and E8-based addressing provide a proposed retrieval structure. Indexes are rebuildable representations of the canonical data, rather than the sole surviving record. A retrieved candidate passes access, version and integrity checks before the application consumes it.

Application integration

Applications call the memory service through a stable API or SDK, receive eligible evidence and pass that evidence to a selected model or deterministic workflow. Deployment can be customer-controlled, private-cloud or embedded, subject to engineering and support readiness. A common evaluation harness measures the integrated result.

Product boundary: the commercial plan funds software, validation and enterprise, operator and OEM integrations from launch. Photonic networks, satellite launch systems and other Trinity Sky infrastructure initiatives are potential future distribution or deployment relationships; their capital requirements and claimed performance are not included in the software and licensing revenue model.

---

08 / Phoenix: turn compression into savings

Phoenix is positioned as a proprietary, lattice-native lossless codec. The reported 130:1 ratio on a raw-data workload is an important development signal. It is a company-reported figure for a specific workload whose full corpus and execution record were not included in the material reviewed for this plan.

What 130:1 means

If 130 units of raw input become one unit of encoded output, the encoded payload is about 0.769% of the original size: a 99.231% raw-byte reduction. That arithmetic applies to the stated input and output boundaries. Headers, dictionaries, codebooks, model weights and any external state must be included or separately amortized in a production benchmark.

No lossless codec can guarantee that reduction for arbitrary input. Already-compressed media, encrypted bytes and high-entropy data must be tested as negative controls. The relevant comparison is the incremental saving against the customer’s current codec and data layout.

Parallel routes to paid adoption

Enterprise storage and asynchronous workloads. Backup, replication and batch movement allow controlled round-trip verification and can tolerate integration before latency-critical deployment.

Operator transport. Qualify stream framing, bounded buffering and decode throughput on links where reduced byte volume can materially improve delivery cost or completion time.

OEM and SDK distribution. The Year 1 plan includes supported implementations licensed to platform partners and covered shipments. Format stability, platform support and long-term decoding are part of the commercial delivery scope. Media and live video need a separate rate-distortion, compatibility and device-power program; raw lossless results do not establish superiority over video standards.

Buyer-facing proof

Each evaluation should produce an invoice-relevant before-and-after result: usable bytes retained or delivered; encode and decode time; CPU and accelerator use; memory footprint; fault recovery; and net annual savings. A license can be priced against verified value, with minimum fees and measurement rules negotiated in advance.

Competitive objective: the best economic result on an identified workload, with exact reconstruction and practical deployment.

---

09 / Memory that preserves the evidence

NEOMORPHIC Memory is designed to keep authoritative knowledge outside transient model context while making it quickly accessible. Its purpose is to reduce repeated rediscovery, preserve the history of changing facts and provide evidence an application can verify.

Canonical records and versions

The technical design uses canonical memory objects, a root describing committed state, a write-ahead log and a version graph. A factual update creates a new authorized version with temporal metadata. Historical records remain distinguishable from current facts, subject to retention and deletion policy. Recovery rebuilds indexes from canonical state and checks consistency before serving queries.

Structured retrieval

FHRR represents relationships through phase-based binding and unbinding. The keyed design describes a finite-group form with modular operations, while E8-based coordinates organize addressing. These structures are proposed mechanisms for efficient association and lookup. Exact reversal of a single binding does not make a superposition of many records immune to interference; occupancy, noise and collisions still require bounds and tests.

Proof-carrying access

A knowledge request is evaluated against a principal, path, authorization epoch and time-limited capability. Candidate acceptance can combine similarity and margin checks, rebinding consistency, membership proofs, authenticated decryption and content hashes. Each check has a specific purpose: authorization, candidate selection, record membership or integrity.

A valid hash proves that returned bytes match a committed object. It does not prove that the object is true, current, relevant to the query or sufficient to support the generated answer. Temporal policy, evidence selection and semantic answer verification remain separate controls.

Practical outcome

The customer receives retrievable evidence with identity, version and provenance. The application can choose the current authorized fact, explain which record it used and decline an answer when evidence is missing or contradictory. That combination is more valuable than speed alone in workflows where an unsupported answer creates expensive rework or operational risk.

Design status: these mechanisms are described in the team’s memory and keyed-access specifications. Scale, adversarial behavior and integrated performance remain subjects for reproducible validation.

---

10 / Addressing model drift and accuracy loss

Model drift is a central reliability problem in this plan because a system that was correct at deployment may fail as its environment changes. The term covers several mechanisms with different remedies. The product must recognize which mechanism is responsible before claiming an improvement.

Define the customer promise narrowly enough to measure

“Zero drift” can describe an invariant for a committed record: the system should not silently change its bytes or identity. It cannot, by itself, describe the future accuracy of every AI answer. The commercial target is to maintain or improve task accuracy as knowledge grows and changes, while detecting failures early enough for controlled correction.

Build a reliability loop into the deployment

Capture the source, version and retrieval path for each evaluated answer. Monitor errors on fixed regression tasks and newly observed cases. When knowledge changes, record the new valid fact and test that the system stops using superseded information. When the task distribution changes, evaluate the model and retrieval policy rather than assuming that more memory will fix the problem.

NIST’s AI Risk Management Framework supports ongoing evaluation and management of AI risks; it is a useful operating framework, not a certification of this product. 17 The proposed service should expose these controls through dashboards and APIs that fit the customer’s existing incident and change-management process.

---

11 / Benchmark evidence and current scope

The available material contains promising internal measurements and later notes that some reruns failed. The figures below preserve their original scope. They do not establish current production performance or a matched comparison with all competing systems.

What can be said today

The historical evidence supports a development thesis around microsecond-scale local lookup and potentially substantial compression on suitable raw data. It does not support an unrestricted claim that the platform is faster or more accurate than every competitor across every benchmark.

The supplied business-plan material references later rerun failures and does not include the complete raw artifacts needed to resolve them. A release-specific replication is therefore part of the seed program. Each published result should carry a code revision, hardware configuration, corpus identifier and evaluation script.

One clock for each claim

Local lookup, batched query throughput, network round-trip time, retrieval-plus-verification and complete answer generation are separate metrics. The plan does not compare a microsecond memory operation to another product’s full API response and call the ratio a system speedup. The commercial benchmark is the customer’s complete task under a matched operating envelope.

---

12 / Validation that wins procurement

The validation program is designed to answer a purchasing question: does the product improve the customer’s workload enough to justify deployment, with predictable behavior when it reaches its limits?

Phoenix protocol

Use representative raw and production-formatted datasets, including already-compressed and encrypted controls. Record data provenance, permitted use, file sizes, redundancy and train/test separation if learned components are involved. Compare with the customer’s current settings and appropriate Zstandard, LZ4 or Brotli configurations; include nvCOMP only when the same GPU environment is available to both sides. 22, 23, 24, 25

Count every byte needed to decode, including framing, dictionaries, shared state and amortized model costs. Verify bit-for-bit reconstruction through hashes across the full dataset. Report compression ratio, encode and decode throughput, p50/p95/p99 latency, peak memory, processor utilization and failure recovery. Test cold starts and warm steady state separately.

Memory and drift protocol

Test retrieval and completed answers separately. Use temporal updates, conflicting records, similar entities, multi-session questions, inaccessible records, deleted records and queries with no valid answer. Adapt LongMemEval-style tasks and customer-specific regression suites without leaking held-out answers into memory preparation. 16

Measure exact-record retrieval, evidence relevance, final-answer accuracy, stale-answer rate, false acceptance, abstention and latency distributions. Report task coverage alongside accuracy so a system cannot appear perfect simply by refusing difficult questions. Increase corpus size, role occupancy, concurrency and update frequency until operating limits are visible.

Release and customer acceptance

Freeze a release candidate, publish a reproducible execution recipe and have an independent evaluator rerun selected workloads. Customer acceptance should require agreed minimum economic benefit and quality, rather than a single universal compression or accuracy threshold. Failed cases remain in the regression set.

Milestone: a benchmark package that a technical buyer can reproduce and map to its own economics. This is the basis for an SLA, commercial comparison and scale claim once the corresponding deployment evidence exists.

---

13 / The customer savings that fund the license

Phoenix creates commercial value by reducing eligible physical bytes. The contract measures savings against the customer’s deployed baseline, including all encoded state and processing costs. The examples below are editable planning cases in Value Economics, not published product benchmarks.

Storage: $30M retained customer benefit

Consider 100 PB of logical information with three physical copies at an avoidable owned-capacity cost of $0.02 per GB-month. The annual capacity cost is $72M. A measured 50% incremental reduction would avoid $36M of that cost. After $3M of annual encode/decode operations and a $3M license, the buyer retains $30M per year.

This example uses physical replicas and a physical capacity cost. A managed storage price that already includes redundancy should not be multiplied by the same replica factor again. Requests, transfers and other service charges remain separate; S3’s pricing structure illustrates these distinct charge categories. 52

Transmission: $175M of value before the license

At 100 eligible EB per year, $0.01 avoidable cost per GB, 25% incremental byte reduction and 80% cash realization, gross realized savings are $200M. Subtracting $0.00025 of codec cost per eligible GB leaves $175M. A 25% share of that benefit produces $43.75M annual Phoenix revenue, leaving the buyer $131.25M.

The same unit economics underpin the original lead model’s 40 EB Year 1 operator scope and $17.5M operator revenue. More contracted, eligible traffic scales revenue; the entire network’s traffic is not automatically licensed or compressible.

What the 130:1 result contributes

Management’s reported raw-data ratio corresponds to a 99.23% payload reduction on that workload. It is the high-upside technical signal motivating customer qualification. The commercial ratio must be measured on the selected workload relative to its existing representation. A blended estate’s benefit depends on the eligible share, achieved reduction and actual cost realization.

The world-scale opportunity is repeated deployment of a favorable unit economic result. It does not require assuming that every global data bill falls by the raw-data compression ratio.

---

14 / Memory efficiency becomes inference capacity

AI inference spends time computing and moving information through memory. Reducing the right data movement can improve throughput, while persistent knowledge can reduce repeated prompt construction and the amount of irrelevant history supplied to the model.

A transparent hardware-performance example

Suppose memory movement accounts for 60% of elapsed time on a measured workload. If an implementation halves that eligible movement and adds overhead equal to 5% of original elapsed time, the new elapsed-time ratio is:

1 − (60% × 50%) + 5% = 75% of baseline time.

That represents a 25% time reduction and approximately 1.33× throughput under unchanged bottlenecks. These are planning calculations, not NEOMORPHIC benchmark results. If compute, interconnects or decoding becomes the limit, the measured outcome will differ.

At one million covered accelerators and $20,000 of avoidable annual cost per accelerator, a 25% useful efficiency improvement represents $5B of annual capacity value. A $600 annual runtime license generates $600M revenue and leaves $4.4B of customer value before other integration expense. A procurement case must establish the avoidable cost and how freed capacity is monetized.

Distinguish the working sets

For a simplified transformer, KV-cache bytes scale with two tensors × layers × KV heads × head dimension × bytes per element × cached tokens × concurrent sequences. A 32-layer example with eight KV heads, head dimension 128, two-byte elements and 32,768 tokens uses about 4 GiB per sequence for KV tensors alone. At 16 concurrent sequences that is 64 GiB before weights and other overhead.

PagedAttention, FlashAttention and quantization address different parts of this problem. NEOMORPHIC must compare against optimized baselines and preserve quality; lossless raw-data compression is not an established KV-cache ratio. 19, 20, 21

Persistent memory reduces repeated work

Retrieve authoritative evidence, enforce access and version rules, and send the relevant subset to the model. Measure retrieval overhead, prompt tokens, time to first token, output latency and cost per correct task. Titans provides another research example of learned long-term memory; it is a comparator, not evidence that NEOMORPHIC has already surpassed it. 58

---

15 / The markets Phoenix touches

Phoenix can participate in large transmission industries by reducing the eligible byte component of customer workloads. Industry revenue is useful context, but includes content, equipment, labor, access services and other spending that a codec cannot capture directly.

Units: nominal US dollars, billions. † 2030 value derived from a publisher’s stated base or endpoint and CAGR; other 2030 values are published forecasts. CAGR periods are the publishers’ periods, not automatically 2025–2030. Rounded values may not reproduce the published CAGR exactly.

Interpret scope before interpreting size

These categories overlap. Satellite internet sits within broader connectivity activity; streaming uses telecom and CDN infrastructure. The mobile-CDN estimate comes from a separately defined study and exceeds the general CDN estimate, illustrating why the labels cannot be treated as a consistent hierarchy or added into one market.

The streaming forecast includes software and delivery activity across categories such as cable, satellite, IPTV and OTT. It is broader than the bandwidth bill of a subscription streaming platform. Similarly, satcom includes equipment and associated services. 3, 4

Research dispersion matters

MarketsandMarkets forecasts a $42.89B CDN market in 2030, compared with TBRC’s $77.94B. Different scopes and assumptions produce materially different estimates. The plan retains the source for each number and uses customer budget discovery to establish the serviceable opportunity. 7, 9

These forecasts support the scale of the industries served. They do not validate an aggregate $4.6T unique addressable market or a Phoenix revenue forecast.

---

16 / Storage and dependable AI expand the opportunity

The revised strategy adds the markets connected to storing data and operating trustworthy knowledge systems. These are substantial existing budget areas even before a dedicated category for persistent AI memory is established.

† Illustrative 2030 extensions using published growth assumptions. Enterprise data management is a broad category encompassing software and services; it is not a standalone estimate for AI memory or drift prevention. 10, 11

The storage entry point

Cloud and enterprise storage buyers pay for capacity, resilience, retrieval and data services. Phoenix targets the portion whose cost falls when fewer bytes must be retained or replicated. Selling through storage platforms can expand distribution, but serviceability depends on data type, decoder integration, customer contracts and vendor incentives.

The memory and reliability entry point

Enterprise AI teams already spend on retrieval, data integration, knowledge quality, observability and software that supports operational decisions. NEOMORPHIC Memory can compete for a defined deployment budget within these categories. A buyer-level opportunity model should use qualified organizations, deployable workloads, annual contract value and expected procurement timing.

IDC’s August 2024 forecast placed worldwide AI spending at $632B in 2028. It is a dated forecast spanning applications, infrastructure and services, included as context for AI investment rather than as a current memory-market estimate. 12

Who should buy first

Prioritize organizations with both costly data operations and a testable knowledge problem: infrastructure operators replicating substantial data; enterprises with changing technical or policy knowledge; research teams preserving decision history; and sovereign or regulated environments needing local control and auditability.

Sizing discipline: establish the eligible workload and the budget owner first. The plan does not add cloud storage, enterprise data management, AI spending and transmission revenues together. A single customer’s spend may appear in several of those totals.

---

17 / The billion-dollar licensing opportunity

Phoenix’s opportunity is the economic value of improving the world’s data infrastructure. Management’s $143B codec-addressable 2030 pool provides a reference for the scale of annual revenue that a broadly adopted compression technology could earn.

These figures use vendor revenue capture as a share of the reference pool. They preserve the market-capture thesis: 0.5–1% represents $715M–$1.43B a year; 5% represents $7.15B. The exact 0.5% calculation is $715M. The $143B denominator is a management sizing assumption, separate from the cited sector forecasts.

A contract model connects the ambition to execution

The lead operating plan targets $115.5M Phoenix revenue in Year 1 and $1.262B in Year 3, generated through enterprise licenses, eligible operator traffic and OEM/SDK agreements. These are three channels for earning revenue from the technology. They are modeled directly rather than by assuming every potential user pays the same fee.

Capture and contract pricing are different measures

A market-capture percentage describes the revenue ultimately earned. Operator contracts may price against net customer savings; enterprise agreements may use annual scope; OEM agreements may combine platform and covered-unit fees. Those contract economics are already included in the operating model. The market-capture table expresses the resulting vendor revenue share.

Expand through supported implementations

An open format or reference core can accelerate interoperability. Optional enterprise management, optimized implementations, support and explicit OEM rights provide paid offerings. The market opportunity grows through useful deployment and distribution. Open-source adoption alone does not automatically create royalties.

The capture table and operating plan are alternative views of the opportunity. They are not added together as separate revenue. Commercial Year 1 is the first 12 months of launch; the 2030 pool remains a separately dated market reference.

---

18 / Competition: compression and data movement

Phoenix competes for a place in an existing data path. Its commercial proposition combines compression efficiency, reconstruction speed and ease of operation. The commercial comparison is against the buyer’s current production system.

Differentiate on the economic result

Incumbent codecs include capable open-source and widely deployed systems. Phoenix cannot rely on a claim that every alternative is old, expensive to license or unavailable. VVC has public encoder and decoder implementations, including Fraunhofer’s VVdeC; implementation availability and commercial adoption are separate questions. 26

Video formats also solve a different optimization problem when they allow controlled loss of visual information. A valid media evaluation needs matched resolution, bitrate, visual quality, decode capability, power consumption and latency. A lossless raw-data result is not evidence of a 130-fold improvement over an already-compressed video stream.

Distribution and interoperability

Start where Trinity Sky can control both encoding and decoding, make installation reversible and provide a long-term decode commitment. Customers need deterministic failure behavior, version compatibility and clear terms for deployment, support and exit. Proprietary licensing may avoid participation in some standard-specific pools, but freedom to operate and third-party rights still require a documented review.

The proposed positioning is measurably better data economics on qualified workloads, supported by reproducible tests and a deployable product.

---

19 / Competition: persistent memory and reliability

The market already contains serious approaches to agent memory and enterprise retrieval. NEOMORPHIC should compete on verifiable behavior and workload economics while integrating with systems customers already use.

The proposed NEOMORPHIC distinction

Combine fast associative retrieval with canonical records, versioned knowledge and proof-carrying access. Measure the result over knowledge updates and extended sessions, then show that the final answer remains supported by the correct evidence. The architecture is attractive only if its additional checks retain useful throughput and predictable operating costs.

A realistic integration strategy

Position the service as a memory and evidence layer usable by multiple models and agent frameworks. Customers may continue to use vector search, graph databases, existing observability and a chosen model provider. Exportable canonical data and rebuildable indexes reduce migration risk and make the product easier to evaluate.

Sovereignty is an important deployment requirement, but other suppliers already offer private and air-gapped environments. NEOMORPHIC’s defensible contribution must be demonstrated at the data and application layers, rather than treating local deployment alone as a unique feature.

---

20 / Five routes into contracted revenue

The commercial architecture supports high-value enterprise scopes and broad distribution from the first commercial year. The launch plan has five recognized revenue lines.

Enterprise licenses and customer value

Year 1 assumes $1M average annual new license values for both SSI/Memory and Phoenix. Annual contracts begin throughout the year, so new licenses contribute 50% of annual value to first-year recognized revenue. The plan targets large data and knowledge estates where the benefit can support that scope.

For Phoenix, a $1M annual fee at a 25% share of net pre-license savings requires at least $4M of annual customer savings. SSI/Memory value can also include fewer costly errors, less rework and more efficient access to knowledge. Each contract defines the paid scope and benefit measurement.

Operator and OEM distribution

Operator pricing uses $0.00175 net savings per eligible GB and a 25% Phoenix share, giving $0.0004375 per GB. The Year 1 plan covers 40 EB of eligible traffic. OEM/SDK revenue combines eight average active platform contracts at $2.5M with 120M covered paid-implementation units at $0.15 each.

One buyer, distinct paid scopes

Fifty SSI and 120 Phoenix enterprise licenses overlap at 25 buyers, producing 145 enterprise parent customers and $14.5M implementation revenue. Operator and OEM workloads are counted separately and exclude already-billed enterprise scopes. An open specification enables interoperability; optional paid implementations and supported enterprise capabilities create the license revenue.

---

21 / MediaTek and Apple: distribution at device scale

MediaTek says its technology powers more than two billion devices each year. Apple reported more than 2.5 billion active devices in January 2026. Annual device reach and an installed base are different licensing opportunities. 35, 36

MediaTek: license the implementation into a platform

The proposed entry point is Phoenix encoding and decoding for eligible device storage, telemetry and edge-AI data flows, followed by qualified integration into selected SoC families. Target the architecture, multimedia, connectivity, AI software and strategic IP licensing teams. The buying case is reduced memory pressure, fewer transferred bytes and more useful work within a device’s power envelope.

The model proposes $0.50 per covered shipped SoC, a $25M creditable annual minimum and a $10M integration program. At one billion covered annual shipments, the running royalty is $500M per year. Those prices and covered volumes are proposal assumptions. MediaTek’s total annual reach supplies the distribution context, not an assertion that every family is eligible.

NVIDIA and MediaTek’s August 2026 collaboration spans custom cloud accelerators, local AI computing and automotive platforms. That provides a concrete ecosystem for proposing platform-level co-design. It is an existing partnership between those companies, not a Trinity Sky partnership. 54

Apple: monetize active memory capability

Propose an annual NEOMORPHIC persistent-memory entitlement on selected devices, plus a separate technical program for eligible compressed storage. Approach silicon architecture, machine-learning frameworks, operating-system performance and licensing teams. The model uses $1 per active covered device per year, reaching 1.5B covered devices before renewal weighting in Year 10.

Apple’s MLX already uses unified memory shared by CPU and GPU. The proposed benefit should therefore focus on working-set size, persistent knowledge and eligible data reduction rather than promising to remove copies that the platform already avoids. 44

Both accounts need measured application performance, battery impact, random-access behavior and long-term compatibility. One paid unit must be counted once across the selected device and component entitlement.

---

22 / NVIDIA and AWS: sell efficiency into the AI factory

NVIDIA: earn a place in the memory path

NVIDIA’s BlueField-4 context-memory platform explicitly addresses the growth of inference context and its movement beyond GPU memory. This is a specific product category into which NEOMORPHIC can propose persistent context management and Phoenix can test eligible encoded data flows. Target the inference software, networking, storage and platform architecture teams. 38

Begin with a CUDA-compatible software evaluation alongside nvCOMP, then test integration with context-management and data-transfer interfaces. Compare complete inference throughput, tail latency, memory footprint and answer quality. A local recall microbenchmark alone cannot establish an accelerator-system speedup. NVIDIA’s published performance claims remain its own competitive baseline. 25, 38

The proposal case uses $100 per new covered accelerator implementation, a $40M creditable annual minimum and $15M integration revenue. Five million annual covered shipments would generate $500M of running royalties. The modeled scope is Phoenix implementation IP, with resale of the same chip excluded from a second royalty.

AWS: annual entitlement across a deployed fleet

Target Annapurna Labs, Neuron, EC2, storage and Bedrock infrastructure teams. AWS’s Trn2.48xlarge specification combines 16 Trainium2 chips, 1.5 TB of HBM and 46 TB/s aggregate HBM bandwidth. These figures show why line-rate decoding and limited metadata overhead are essential for a memory-path proposal. 42

The first commercial product can sit in the software/runtime layer: persistent context, retrieval, checkpoint movement and eligible storage streams. Future silicon integration requires a separate engineering and qualification program. AWS and NVIDIA’s announced plans include custom memory collaboration as well as large fleet expansion. 55

The proposed AWS license is $600 per active covered accelerator per year, with a $50M annual floor and $10M integration scope. One million covered accelerators creates a $600M annual license. The entitlement covers NEOMORPHIC runtime functionality; the model excludes duplicate fees for Phoenix rights already paid by a chip supplier or an operator agreement.

---

23 / Google and Microsoft: recurring cloud deployment

Google: TPU, cloud memory and sovereign infrastructure

Google’s TPU documentation describes model parameters in high-bandwidth memory feeding the compute units. This establishes a clear technical target: reduce eligible memory or storage traffic while preserving the execution path’s required throughput. 43

Approach TPU architecture, Google Cloud infrastructure and enterprise AI platform teams with a reproducible workload package. The first integration proposal is a supported persistent-memory service and optimized runtime adapter; a future compression block in a controller is a separate design program. Google Distributed Cloud’s air-gapped offering also provides a documented route for evaluating sovereign deployments. 31

The proposal model uses $600 per active covered accelerator per year, a $40M creditable annual minimum and a $10M integration scope. The modeled expansion reaches 9M covered units before renewal weighting in Year 10. That is a scenario assumption, not a disclosed Google fleet count.

Microsoft: Azure runtime and enterprise memory

Target Azure infrastructure, silicon systems, enterprise AI and partner engineering teams. The initial product should connect persistent enterprise evidence to AI workflows and test whether eligible compressed representations improve memory or storage efficiency. Azure deployment can create a recurring infrastructure license; Windows and OEM device distribution would require separately defined commercial rights.

Microsoft reported more than $100B in annual Azure revenue in FY2026. Its scale supports a large purchasing opportunity, while a sale still depends on quantified customer value and technical integration. 40

The proposal model applies the same $600 active-accelerator annual fee, $40M minimum and $10M integration scope to an illustrative Azure cohort. It reaches 10M covered units before renewal weighting in Year 10. Procurement, performance and security acceptance precede fleet rollout.

Sell to the platform and the workload owner

An architecture team validates feasibility; the service owner validates cost per successful request; procurement negotiates scope and price. The account plan must satisfy all three. Cloud marketplace distribution can extend reach, but marketplace listings alone are not platform design wins or committed revenue.

---

24 / Intel, IBM and the memory supply chain

Intel: software proof, then hardware IP

Intel’s Query Processing Library already supports accelerated compression and analytics using the Intel In-Memory Analytics Accelerator, with software fallback. This is both a competitive baseline and a practical customer discussion point. Phoenix should first demonstrate incremental value through software integration; an existing accelerator cannot be assumed to decode a new format without hardware support. 45

Target Xeon platform architecture, accelerator software and IP ecosystem teams. The proposal case uses $2 per covered processor or controller shipment, a $20M annual minimum and $10M integration scope. At 250M modeled annual covered shipments, the unweighted running royalty is $500M. Volumes are assumptions for a multigeneration program, not an Intel shipment forecast.

IBM: storage-controller and enterprise deployment

IBM FlashSystem already incorporates advanced storage and data-reduction capabilities. Propose a controlled comparison on eligible data, including usable capacity, tail latency, recovery, replication and total controller cost. The buyer needs improvement over its installed baseline. Target IBM Storage engineering, product management and enterprise software partnerships. 46

The model uses $100 annually per active covered controller, a $5M minimum and a $5M integration program. A controller entitlement covers a defined supported function; it does not imply a royalty on every drive or record attached to it. IBM’s infrastructure and enterprise distribution also creates a route to larger software deployments. 47

Memory vendors and design-IP channels

Micron’s HBM portfolio illustrates the bandwidth and capacity demands of AI systems. Micron also documents a design-IP partner network, including Cadence and Synopsys. These are specific routes for proposing a qualified block or implementation to component and platform designers. 48, 49

The technology would generally integrate through a controller, logic layer, firmware or runtime. A passive memory device does not execute a new codec merely because it stores data. The separate model cohort covers additional, non-overlapping memory/controller designs at $0.20 per unit. Samsung and SK hynix are additional prospective memory accounts for qualification, without an assumed agreement or supplied benchmark on their products.

---

25 / A billion-dollar annual sovereign business

Governments need to retain trusted records, operate secure AI systems and move data across constrained infrastructure. A national or agency license can combine Phoenix data efficiency with NEOMORPHIC memory, access control, versioning and auditable evidence for civilian administration, research, logistics and maintenance.

Large software procurement has a precedent

The U.S. Army’s 2025 Palantir arrangement permits purchases of up to $10B over ten years. It is a procurement ceiling and does not commit new purchases. That distinction matters: the NEOMORPHIC model earns revenue only from funded annual scopes. 57

AWS and NVIDIA announced plans for secure federal AI infrastructure including 100,000 GPUs. This identifies an existing platform channel for proposing licensed infrastructure capability. 55

Build an annual portfolio, program by program

An illustrative $100M annual program could comprise $40M of national data-platform entitlement, $30M of secure AI-memory runtime, $20M of distributed-site entitlement and $10M of support and operations. Each component needs a defined deployment and acceptance schedule. Hardware, cloud bills and prime-contractor pass-through are excluded from NEOMORPHIC license revenue.

Ten funded programs at $100M each create $1B annual revenue; forty create $4B before renewal weighting. The model begins sovereign deployment in Month 25 and uses 98% annual renewal weighting, producing approximately $960M in Year 5 and $3.47B in Year 10. These are proposed portfolio economics, not awarded government contracts.

The route to an award

Target national digital-infrastructure agencies, defense enterprise IT, research laboratories and sovereign cloud operators. Start with a paid, bounded workload evaluation; move to an authorized environment through the agency or an established cloud/prime channel; negotiate funded task orders and annual expansion options. Planning gates cover data residency, air-gapped operation, accreditation, supply-chain documentation and local support.

For U.S. commercial software, DFARS policy recognizes customary commercial licenses subject to federal requirements and negotiated rights. Background IP, government-funded changes and disclosure obligations must be defined in the agreement. This offers a route to recurring commercial licensing while preserving the underlying IP under negotiated terms. 53

---

26 / Structure the license for recurring IP income

The business model licenses technology usage and support over defined terms. Commercial documents should call this a term technology license where appropriate; “leasing IP” describes the economic intent without implying a particular accounting classification.

These are proposed commercial assumptions for negotiation, not observed NEOMORPHIC prices. Each license must leave the buyer an attractive share of measured benefits.

Minimums, renewals and usage reports

Use a three- to five-year commercial framework with annual minimums, defined renewal rights and a schedule for additional product generations. Credit the minimum against running royalties: annual license revenue is the greater of the earned royalty and the floor, subject to actual contract terms. The workbook bills annual floors upfront and excess royalties monthly; the payment lag is editable.

Rights that scale with distribution

Define the licensed binary, SDK, firmware or hardware design; allowed subsidiaries and subcontractors; reporting of shipped or active units; authorized fields of use; derivative improvements; compatibility obligations; support; and generation refresh. An OEM-paid component should not trigger a second identical royalty when resold by a cloud provider.

Revenue and cash are different

The model separately schedules earned licenses, integration revenue, invoices, collections, receivables and deferred revenue. Upfront cash does not itself establish earned revenue. Final recognition depends on the executed contract’s performance obligations; IFRS 15 provides the relevant revenue-recognition framework. 56

---

27 / Commercialize, distribute and scale

The go-to-market plan targets paid commercialization of the supported core from launch, with enterprise selling, operator agreements and OEM integration running in parallel. Initial design partners improve references and pricing proof while the wider commercial pipeline advances.

Resource the sales target

Year 1 requires 181 commercial relationships: 145 enterprise parents plus 20 operator and 16 OEM year-end agreement equivalents. At 25% qualified-opportunity conversion, that requires approximately 724 qualified opportunities. The operating plan funds 28.8 productive seller equivalents and 48 deployment engineers, including direct selling and supported distribution.

Budgets match the ambition

The original $3M seed GTM allocation is $1M sales/BD plus $2M pilots. The accelerated operating plan budgets $18M Year 1 sales and marketing plus $2M co-funded pilots, supported by seed and customer receipts. The seed allocation is included in financing; it is not added again as an expense.

Pilot funding remains $500K financial, $400K scientific, $400K engineering and $700K defense/sovereign. Across these sectors, evaluations connect storage, transmission and knowledge reliability to a production purchasing decision.

Commercial status and measurement

The company presentation reports prototype and design-partner activity. 32 The account counts above are execution targets, not an assertion that contracts are already signed. Track earned revenue, billings, collections, implementation capacity, customer benefit and renewal separately.

---

28 / Execute from the first commercial month

NEOMORPHIC’s launch premise is to sell supported licensable implementations while broader platform development continues. Validation, product engineering and commercial delivery proceed together; future hardware or network programs do not define the start date of the software business.

Months 1–3: launch the supported core

Finalize enterprise and partner scopes, rights, deployment packages and service commitments. Activate the initial design partners and larger qualified pipeline. Begin annual license invoicing and implementation. Reproduce release-specific compression and memory measurements on representative workloads, and attach the operating envelope to each supported product.

Months 4–12: scale commercial delivery

Target 50 SSI and 120 Phoenix enterprise license wins, with 145 parent buyers after overlap. Deliver operator and OEM programs corresponding to 20 and 16 year-end agreement equivalents. The Year 1 plan targets $155M revenue, $53.7M EBITDA and $39.3M net income. Monthly EBITDA first turns positive in Month 4 in the original model.

Year 2: compound retained accounts and distribution

Target $611.7M company revenue, including $473.1M from Phoenix. Add 100 SSI and 250 Phoenix enterprise licenses, increase eligible operator traffic to 200 EB/year and expand covered SDK shipments to 400M. Operating expense grows with funded selling, engineering and customer support capacity.

Year 3: pass the billion-dollar revenue milestone

Target $1.605B company revenue and $617.2M net income, including $1.262B Phoenix revenue. New Phoenix enterprise scope averages $1.5M annually. The plan includes 500 eligible operator EB/year and 900M covered SDK units. A prospective $250M financing supports expansion and strategic flexibility; it is not a revenue line.

Product and operating discipline

Maintain stable formats, supported decoders, controlled knowledge updates, recovery and regression testing. Direct resources toward the workloads and channels demonstrating the strongest customer economics. The operating scorecard connects technical performance to contracts, delivered activity, renewal and collected cash.

---

29 / Organization, IP and delivery

Trinity Sky presents NEOMORPHIC through Trinity Global Partners LLC (EGD33), with Palm Beach, Florida as its operating location. Entity structure, contracting authority and ownership of licensed technology should be documented for financing and customer diligence rather than inferred from brand names.

Leadership and near-term roles

The current presentation identifies Enzo Garoche and Skyler Trotter as founders, and Peter Dwight Sahagen as president. 32 The operating plan requires clear ownership of codec engineering, memory systems, independent evaluation, enterprise implementation and commercial execution. Headcount and compensation should be set within the approved budget after the required capabilities and hiring sequence are confirmed.

Product delivery model

Maintain one supported release train with regression suites for both products. Separate research builds from customer-supported releases. Publish decoder compatibility and data export policies before customers depend on proprietary formats. Track onboarding time, support load and incident resolution to ensure that services effort does not consume the assumed software margin.

Intellectual property

Build an invention and ownership register covering Phoenix, memory representations, addressing, access protocols and implementation methods. Map authorship, assignments, filing status, third-party licenses and intended fields of use. Select patent protection, trade-secret treatment and selective disclosure based on the actual invention and commercial distribution model.

Patents filed or associated with other Trinity Sky initiatives do not automatically protect every claim in this plan. The company should make any patent count or exclusivity representation only against a verified portfolio schedule. Licensing without a standards pool does not eliminate the need to review third-party rights.

Data and operational control

Customer records require explicit retention, deletion, backup, encryption and access rules. Immutable history and lawful deletion requirements must be reconciled through the storage and key-management design. Integrity checks, audit trails and deployment isolation support customer control; they do not confer automatic regulatory certification.

The company’s commercial strength will depend on its ability to make advanced engineering dependable enough to procure, deploy and operate.

---

30 / Revenue and profitability from Year 1

The management lead case targets more than $100M annual revenue in Year 1 and more than $1B in Year 3, with positive net income in the first commercial year. The original integrated financial model supplies the contract, cost, billing and cash schedules.

Revenue recognized from delivered commercial scope

Year 1 recognizes $25M SSI/Memory licenses, $60M Phoenix enterprise licenses, $17.5M operator revenue, $38M OEM/SDK revenue and $14.5M implementation. The model assumes annual enterprise contracts start throughout the year and contribute half of their annual value in the first year.

Phoenix retention uses 5% annual logo churn and 8% expansion among retained accounts; SSI uses 5% churn and 10% expansion. These produce 102.6% and 104.5% modeled net revenue retention. Operator usage and covered-unit fees follow earned activity.

Profitability is included in the operating build

Year 1 gross margin is 72.7% and EBITDA margin 34.6%. The model includes $59M operating expense, $6M capex and $42.35M direct costs. Five-year depreciation and a simplified 25% tax rate on positive EBIT produce $39.34M Year 1 net income. EBITDA is distinct from operating cash flow.

Timing and basis: Year 1 is the first 12 commercial months. These are management-target projections with explicit assumptions; revenue, annual run rate, billings and cash receipts remain separate.

---

31 / A path to multi-billion-dollar revenue

The management plan scales enterprise cohorts, operator traffic and OEM distribution. The larger distribution case is shown alongside the lead target; slower adoption and isolated stresses remain separate analytical views in the workbook.

The PDF chart plots Years 1–5 for company revenue, Phoenix revenue and the expanded company scenario. The table provides the full ten-year values.

The next stages of scale

The lead case targets $5.05B company revenue in Year 5, including $3.96B Phoenix revenue. The Year 10 extension reaches $25.29B company revenue under the disclosed account, price, distribution and cost assumptions. These longer-range outputs are an operating scenario, not a claim that the corresponding market share has already been won.

Expanded distribution uses 50% more new accounts, eligible traffic and covered units; 15% higher new enterprise and platform fees; and 60% higher operating expense and capex. It reaches $2.68B company revenue in Year 3. No probability is assigned to either case.

---

32 / Capital and cash support the launch

The $20M initial seed funds the start of the commercial program. The larger operating budget is supported by modeled customer receipts, so the seed allocation is not the ceiling on Year 1 delivery or growth.

The commercial operating budget

Year 1 budgets $59M operating expense, $6M capex and $42.35M direct delivery costs. Operating expense includes $15M R&D, $12M enterprise engineering, $4M validation, $2M pilots, $3M security/IP, $18M sales/marketing and $5M G&A. Initial seed allocations are not counted again as additional expenses.

Advance billing and customer collections

The original monthly model assumes annual enterprise invoices upfront, 45-day collections and 30-day direct-cost payments. Year 1 produces $240M billings, $204.03M customer collections and $109.55M closing cash. The minimum month-end cash balance over the first 24 months is $9.08M, starting with the $20M seed.

A separate 90-day receipt-delay stress requires approximately $18.49M additional cash to maintain a $5M buffer. The lead case keeps its original timing; the stress identifies a working-capital requirement rather than replacing the revenue target.

Expansion financing

The original $250M Year 3 Series A target remains an expansion option. Under the lead receipt assumptions, operations are profitable and cash-generative before that financing. The workbook also shows cash without the round. Neither financing target represents committed capital.

---

33 / The 120-month global-IP expansion case

The original management plan remains the lead commercial commitment for planning: $155M in Year 1 and $1.605B in Year 3. A separate global-IP case shows the larger opportunity from hardware distribution, deployed runtimes and sovereign programs.

The expansion model contains 120 monthly periods and eleven separately scheduled licensing cohorts. From Month 13, it removes the original OEM platform and SDK line and substitutes the new global-IP schedules. Original enterprise, operator and service revenue continues on distinct paid scopes. The two cases are alternatives, not revenue added together.

A defined path to the larger numbers

Each cohort has an activation month, non-recurring engineering (NRE) delivery term, annual minimum, covered-unit price, annual volume schedule, renewal rate and direct-cost rate. The model adds a dedicated hardware/IP operating program, capital expenditure, invoicing and collection schedules. All named accounts are proposed targets; there are no assumed signed contracts represented as actual results.

The Year 10 case earns approximately $23.89B from the global-IP cohorts, including $3.47B in sovereign license revenue. Combined with the retained core scopes, total modeled revenue reaches $47.34B. This is the economic potential of broad multigeneration adoption, with the required coverage exposed in the workbook.

---

34 / How platform licensing compounds

The PDF chart compares original lead and global-IP company revenue across ten years. The following table gives the values; IP Annual contains the underlying workbook data.

One accepted implementation can generate successive years of licensed shipments, deployed runtime use and supported product generations. The global-IP case models these mechanisms separately and replaces the original OEM/SDK line from Year 2.

Device distribution and recurring infrastructure

Shipped-component royalties follow product adoption. Active-device and accelerator licenses renew over the installed covered fleet. Additional product families expand the paid unit base. The proposed minimums and implementation fees support integration, while reported usage controls earned royalties.

A sovereign portfolio contributes billions annually

The expansion case recognizes approximately $300M in sovereign licenses in Year 3, $960M in Year 5 and $3.47B in Year 10. The annual volume assumptions reach forty $100M programs before renewal weighting. Each represents a defined, funded scope; a multiyear purchasing ceiling is not annual revenue.

The graph and table link to the workbook’s annual summaries, which reconcile to all 120 monthly periods. Coverage, pricing, activation dates, renewals and expenses remain editable. These are adoption scenarios rather than a valuation or a statement of signed pipeline.

---

35 / What drives the global licensing revenue

These are editable adoption assumptions, not disclosed customer fleets or shipment forecasts. Annual renewal weighting is 98%; volumes are conditional on retained contracts. The forecast’s active-unit inputs represent annual average covered installations. Shipped-unit inputs represent newly covered annual shipments. Both are prorated for the cohort’s activation date.

Convert a target into an earned royalty

The buyer first accepts the workload result and licensed scope. Software or firmware deployment can then expand across approved units. New silicon integration follows architecture acceptance, implementation, verification and product qualification. The Month 13–25 entries assume initial supported software/platform deployments; new hardware designs require their own release gates.

The model credits annual minimums against running royalties. Integration fees are separate delivered services. Additional memory and edge cohorts exclude units already paid under the named manufacturer cohorts. Cloud runtime and chip royalties cover distinct functionality; identical rights are charged once.

---

36 / Monthly cash, delivery cost and capital

The model provides separate Lead 120 Months and IP 120 Months schedules, with annual reconciliation and a balance-sheet check in every expansion month. Commercial Month 1 is the first month of launch; January 2027 is an editable calendar reference.

Preserve the original commercial targets

The lead case includes $20M seed, $59M Year 1 opex, $6M capex and $42.35M direct costs. It targets $240M billings, $204.03M customer collections and $39.34M net income in Year 1. Annual enterprise invoices and 45-day collections support the working-capital profile. A $250M Year 3 financing option remains separate from revenue.

Fund the global-IP delivery program

The expansion adds R&D, sales/channel and IP/security/G&A spending of $60M in Year 2, $120M in Year 3 and $1.10B in Year 10. Incremental capex rises from $10M in Year 2 to $150M in Year 10. Direct license costs range from 10–18% for hardware and cloud cohorts and 35% for sovereign programs; integration delivery costs are modeled at 70% of integration revenue.

These assumptions fund an asset-light licensing organization and its software, qualification and support work. They do not include building semiconductor fabs, purchasing customer fleets or financing national infrastructure construction. Such obligations would require a separate capital plan.

Cash follows the executed terms

The new license schedules invoice annual floors and integration fees upfront, earn them over their service periods and bill excess royalties monthly. Collections lag new-IP invoices by two months; direct-cost payments lag by one. Annual tax expense uses a simplified 25% rate on positive EBIT, with quarterly cash payments. Capex is depreciated over 60 months for the added IP program.

Under the expansion assumptions, Year 3 ends with approximately $3.81B cash and Year 10 with $101.89B. These undistributed balances reflect no dividends, acquisitions or buybacks. They are not additional revenue or a valuation. The workbook exposes price, coverage and activation-delay controls so readers can assess the collection and adoption requirements directly. A 20% unit-price reduction yields $4.45B Year 3 and $42.56B Year 10 revenue; a twelve-month activation and volume delay yields $2.52B and $43.95B, with other inputs unchanged. Annual minimums limit the effect of lower covered usage.

---

37 / Win a workload. Qualify a platform. Expand the fleet.

Months 1–6: commercial proof and paid integration

Execute the original enterprise and operator launch plan while forming named platform account teams. Build one customer-owned benchmark package for storage, transmission and persistent AI memory, with exact reconstruction, current-release identifiers and cost-per-task comparisons. Negotiate paid evaluations with explicit success criteria and access to representative workloads.

Months 7–18: supported software and design commitments

Deliver production SDKs, runtime adapters, telemetry and compatibility support. Quantify unit economics on customer hardware, then negotiate minimum license commitments, audit mechanisms and product-family rights. Start hardware feasibility work with architecture teams: synthesis, power, performance, area, worst-case latency and memory overhead.

Months 19–36: broader device and cloud entitlement

Expand accepted deployments across covered platforms. The new model activates device cohorts from Month 19, Intel and additional edge designs from Month 25, and additional memory-IP designs from Month 31. These dates are commercial assumptions; each hardware release must satisfy its own qualification gate. Sovereign programs enter from Month 25 through funded scopes and authorized environments.

Months 37–120: successive generations and territories

Renew supported runtimes, expand eligible fleets and add paid product generations. Build regional deployment capacity and authorized partners. Monitor customer concentration, discounting, technical substitution and the share of royalties tied to a single platform. Renewal rights should maintain value through software and hardware refreshes.

The account review that controls expansion

Track benchmark acceptance; paid evaluation value; implementation milestones; signed minimums; eligible shipped or active units; net price; recognized revenue; collections; support cost; and retained customer savings. A target account becomes a forecasted contract only when the internal sales stage and evidence justify it.

The ambition is worldwide device and infrastructure distribution. The operating mechanism is repeated, measurable acceptance by platform owners, with unit-based licensing that preserves customer economics and scales across future generations.

---

38 / Risks, responses and operating measures

The plan targets rapid commercialization at substantial scale. Management should track the commercial and technical variables that determine whether the contract, distribution and profitability targets are achieved.

Monthly executive scorecard

Report qualified pipeline by workload; paid pilots started and completed; production conversions; annual recurring revenue; gross margin by product; implementation time; support load; and cash against the monthly commercial schedule. Do not count expressions of interest, free credits or unexecuted partnerships as sales.

The technical scorecard should show current-release results, regressions and boundaries. Include compression economics, retrieval latency distributions, evidence accuracy, knowledge-update behavior and fault recovery. Historical internal numbers remain labeled until the new release reproduces them.

The operating milestone

An attractive next financing is built on a repeatable chain: a costly customer problem, a reproducible technical improvement, a clear deployment path, an executed contract and evidence of renewal or expansion. This is the operating test of the NEOMORPHIC thesis.

---

39 / Market and financial alignment

Management’s commercial-launch targets anchor the integrated financial model. Updated third-party forecasts define the market context, while contract volumes, license scope and distribution agreements build the company’s growth plan.

Market updates preserve scope and dates

Telecom uses the cited $1.90T 2025 base and $2.46T 2030 forecast. Ericsson’s 515 EB/month estimate refers to 2031 including fixed wireless access. 1, 2 Sector estimates overlap and remain individually identifiable.

Derived 2030 values use the publisher’s stated assumptions: streaming $969.56B × 1.17^4 = $1,816.85B; satellite internet $38.22B ÷ 1.1785 = $32.43B; cloud storage $197.80B × 1.193^4 = $400.67B; enterprise data management $137.30B × 1.128^4 = $222.28B. 3, 5, 10, 11

---

40 / Model and evidence traceability

The business plan and six companion documents use the same management lead case. The workbook integrates the original financial schedules with nine new licensing, monthly, economics and research schedules in a coordinated 29-sheet model.

Commercial source of truth

The lead case is the management-target model revised on 5 September 2026: $155M company revenue in Year 1, $611.685M in Year 2 and $1,604.884M in Year 3. Phoenix revenue is $115.5M, $473.06M and $1,262.371M. Year 1 is the first 12 commercial months. These targets and their contract assumptions are retained in this edition.

The workbook’s Inputs Scale, Model Scale and Cash Scale sheets contain the lead inputs, integrated forecast and monthly billing/cash model. Summary presents the lead case first. Expanded and delayed cases are separate. Sources preserves the original model references; Markets 2030 supplies the newer sector estimates cited in this document.

Financial control

Recognized revenue, annual license run rate, billings, receivables, deferred revenue, EBITDA, net income and cash are distinct outputs. The integrated balance sheet reconciles. The $143B capture table is not added to operating revenue; its percentages describe vendor revenue share. Operator savings-based pricing is already included once in the contract model.

Technical evidence

Historical internal benchmark identifiers include fhrr-recall-20260612T215905Z.json, fhrr-phase7-gates.json, fhrr-suite-B1-20260616T204023Z.json and snn_benchmark_report.json. The 130:1 Phoenix figure is management-reported. Their scope is preserved in the evidence section; current-release replication remains separate from financial scenario design.

The team’s proof-carrying knowledge and keyed-access specifications describe canonical storage, retrieval and acceptance checks. Numbered references identify external forecasts, primary technical papers and official product documentation. The current company presentation supplies the stated development and pipeline context. 32

Companion presentation: 3sky.ai/deck

Expanded model: License Inputs, License Months, IP 120 Months and IP Annual model eleven licensing cohorts. Lead 120 Months presents the original monthly P&L and cash. Global Spending, Licensing Sources and Value Economics connect the market evidence to specific customer economics. The expanded source register contains 58 references.

---

Works cited / Markets and demand

1 Ericsson. Mobile network data traffic forecast. Current forecast, accessed 8 September 2026.

2 Mordor Intelligence. Telecom Services Market. 27 June 2025; 2025–2030 forecast.

3 Fortune Business Insights. Video Streaming Market. 24 August 2026; 2026–2034 forecast.

4 The Business Research Company. Satellite Communication Global Market Report. 2026 edition; accessed 8 September 2026.

5 Mordor Intelligence. Satellite Internet Market. 2026–2031 forecast; accessed 8 September 2026.

6 The Business Research Company. 5G Satellite Communication Global Market Report. 2026 edition; accessed 8 September 2026.

7 The Business Research Company. Content Delivery Network Global Market Report. September 2026; accessed 8 September 2026.

8 The Business Research Company. Mobile Content Delivery Network Global Market Report. 2026 edition; accessed 8 September 2026.

9 MarketsandMarkets. Content Delivery Network Market. 2025–2030 forecast; accessed 8 September 2026.

10 Fortune Business Insights. Cloud Storage Market. 2026–2034 forecast; accessed 8 September 2026.

11 Grand View Research. Enterprise Data Management Market. 2026–2033 forecast; accessed 8 September 2026.

12 IDC. Worldwide Spending on Artificial Intelligence Forecast to Reach $632 Billion in 2028. 19 August 2024; historical forecast vintage.

13 International Energy Agency. Energy and AI: Energy demand from AI. 2025.

---

Works cited / AI reliability and memory

14 Modarressi et al.. NoLiMa: Long-Context Evaluation Beyond Literal Matching. ICML 2025; arXiv:2502.05167.

15 Hsieh et al.. RULER: What’s the Real Context Size of Your Long-Context Language Models?. 2024; arXiv:2404.06654.

16 Wu et al.. LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. ICLR 2025; arXiv:2410.10813.

17 NIST. Artificial Intelligence Risk Management Framework (AI RMF 1.0). January 2023; NIST AI 100-1.

18 Liu et al.. Lost in the Middle: How Language Models Use Long Contexts. 2023; arXiv:2307.03172.

19 Kwon et al.. Efficient Memory Management for Large Language Model Serving with PagedAttention. 2023; arXiv:2309.06180.

20 Dao et al.. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. 2022; arXiv:2205.14135.

21 Zandieh et al.. TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate. 2025; arXiv:2504.19874.

---

Works cited / Competition and company

22 Meta / Zstandard. Zstandard official project. Accessed 8 September 2026.

23 LZ4 project. LZ4: Extremely Fast Compression Algorithm. Accessed 8 September 2026.

24 Google. Brotli official repository. Accessed 8 September 2026.

25 NVIDIA. nvCOMP documentation. Accessed 8 September 2026.

26 Fraunhofer HHI. VVdeC: Versatile Video Coding Decoder. Official repository; accessed 8 September 2026.

27 Chhikara et al.. Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory. 2025; arXiv:2504.19413.

28 Rasmussen et al.. Zep: A Temporal Knowledge Graph Architecture for Agent Memory. 2025; arXiv:2501.13956.

29 Letta. Stateful agents documentation. Accessed 8 September 2026.

30 Pinecone. Indexing overview. Accessed 8 September 2026.

31 Google Cloud. Google Distributed Cloud air-gapped documentation. Accessed 8 September 2026.

32 Trinity Sky. NEOMORPHIC SSI investor deck. Company presentation; accessed 8 September 2026.

---

Works cited / Global IP and customer research 1

33 Gartner. Worldwide IT spending forecast: $6.31 trillion in 2026. 22 April 2026.

34 Gartner. Semiconductor revenue forecast to exceed $1.3 trillion in 2026. 8 April 2026.

35 Apple. Apple reports first quarter results. 29 January 2026.

36 MediaTek. Company overview: more than two billion devices every year. Accessed 8 September 2026.

37 NVIDIA. Fourth quarter and fiscal 2026 financial results. 25 February 2026.

38 NVIDIA. BlueField-4 inference context memory storage infrastructure. 5 January 2026.

39 Groq. Non-exclusive inference technology licensing agreement with NVIDIA. 24 December 2025.

40 Microsoft. FY2026 fourth-quarter and annual results. 29 July 2026.

41 Amazon. Fourth quarter and full-year 2025 results. 5 February 2026.

---

Works cited / Global IP and customer research 2

42 AWS. Amazon EC2 Trn2 instances: technical specifications. Accessed 8 September 2026.

43 Google Cloud. Cloud TPU system architecture. Accessed 8 September 2026.

44 Apple ML Research. MLX: array framework for Apple silicon. Official repository; accessed 8 September 2026.

45 Intel. Intel Query Processing Library (QPL). Official repository; accessed 8 September 2026.

46 IBM. FlashSystem storage portfolio. Accessed 8 September 2026.

47 IBM. Fourth quarter and full-year 2025 results. 28 January 2026.

48 Micron. High bandwidth memory portfolio. Accessed 8 September 2026.

49 Micron. Design IP partner network. Accessed 8 September 2026.

50 Arm. Fourth quarter and fiscal year 2026 results. 6 May 2026.

---

Works cited / Global IP and customer research 3

51 International Energy Agency. Energy and AI: executive summary. 2025.

52 AWS. Amazon S3 pricing and charge categories. Accessed 8 September 2026.

53 Acquisition.gov. DFARS 227.7202-1: commercial software licensing policy. Effective 7 May 2026.

54 NVIDIA and MediaTek. Expanded edge-to-cloud AI computing partnership. 31 August 2026.

55 AWS and NVIDIA. Two million additional GPUs and federal AI infrastructure. 26 August 2026; GPU deployment planned for 2027–2028.

56 IFRS Foundation. IFRS 15: Revenue from Contracts with Customers. Accessed 8 September 2026.

57 Reuters. US Army pools contracts into up to $10 billion Palantir deal. 31 July 2025.

58 Behrouz, Zhong and Mirrokni. Titans: Learning to Memorize at Test Time. 2024/2025; arXiv:2501.00663.