The Hidden Lock-In Beneath Inference and Physical AI.
Why substrate selection in 2026 is a five-to-twelve-year capital commitment, and what the board has to govern.
By Armando Pereira | Founder, PVentures Consulting | Senior Member IEEE | Co-founder, OpenFog Consortium (IEEE 1934) | President, Autonomous Vehicle Computing Consortium | Former VP/GM Optical BU, Centillium Communications (CMOS PON SoC, NTT-qualified)
Current, specific, not obvious.🎯
🧭 Welcome back to The Sextant™
The Sextant™ is a recurring series of board-level briefings on the hidden constraints shaping the infrastructure in deep-tech. Each issue isolates one constraint and ends with the questions a director should carry into the next board meeting.
This is the second of two inaugural Sextant™ issues shared in full, complimentary, with the entire Fog Signal reader base. Issue one, 6 May 2026, named capacity as the first hidden constraint in AI compute. From issue three onward, the series is available only with a paid subscription.
1️⃣ Executive Summary
Two weeks ago, this series named the first hidden constraint: packaging, not silicon. Today, we name the second, and it shows up in two domains at once. In inference, the workload has separated from training, and the substrate is fracturing into a hyperscaler core and a sovereign edge. In physical AI, three Tier 1 OEMs signed commercial humanoid contracts in H1 2026. Both converge on one board question: who selects the substrate, on what evidence, and against what governance trigger.
What is happening?
Anthropic committed $200 billion over five years to 5 GW of Google Cloud TPU capacity; Anthropic and OpenAI now account for more than half of the major cloud providers’ $2 trillion revenue backlog. BMW Spartanburg ran a Figure 02 humanoid through 30,000 X3 vehicles at 99 percent placement accuracy. Mercedes funded Apptronik at a $5 billion valuation. Toyota signed with Agility Robotics at roughly thirty dollars an hour.
Why does it matter now?
Inference substrate decisions taken in 2026 carry compute architectures through 2031 to 2033. On-robot substrate commitments run 60 to 120 months. The chassis on the floor and the GPU in the rack are replaceable within 24 months. The substrate underneath them is not.
What breaks if ignored?
RFPs negotiate the visible layer; substitutability costs lives in the invisible layers. Capital allocation runs blind through the substrate decision and arrives at the audit committee as a sunk position.
What are the winners doing?
Substrate qualification gates before commitment, second-source mandates on the foundation-model layer, four-layer RFP discipline, and a dashboard that names the substrate at each layer for every active deployment.
Board-level actions for the next 90 days are detailed in Section 7: Strategic Options.
Compress complexity into decision clarity.💎
2️⃣ The Shift: From Procurement Decision to Capital Commitment
The mental model that carried enterprise IT through the 2010s treated compute as a vendor procurement question. The chip was a BOM line; the cloud was a contract; the model was an API. Every layer was renegotiable inside a planning cycle. That model is breaking in two places at once.
In inference, the workload has separated from training at the silicon layer. Broadcom reported 45% custom ASIC growth against 16% for the GPU segment in the same quarter. Midjourney cut inference costs by 65% by moving to Google TPUs. NVIDIA responded with NVLink Fusion, a fabric absorption play that participates in revenue regardless of which silicon handles the tokens. The silicon decision is no longer a vendor selection; it is a capex agenda item on a 5-to-7-year commitment cycle.
In physical AI, the same shift moves through the factory floor. NVIDIA Jetson Thor with IGX Thor carries a 10-year lifecycle. Qualcomm Dragonwing IQ10 delivers 700 TOPS, sparse with 30% better power. The chassis is replaceable every 12 to 24 months; the on-robot substrate runs 60 to 120 months. The foundation-model layer locks 24 to 36 months of integration once trained against a behavior envelope.
The standards body record has seen this pattern. OpenFog and IEEE 1934 followed exactly this trajectory: contribute the reference architecture as a public standard, then sell the implementation underneath. NVIDIA is running the same play with Isaac GR00T. The chassis is the cell phone; the substrate underneath is the carrier.
What I thought was the constraint is not. 🔄
3️⃣ The Bottleneck: The Board Has No 4-Layer Instrument
The bottleneck is institutional, not technical. Most boards do not receive a substrate reporting line on either AI deployment or physical-AI deployment. Three named instruments are missing.
The substrate dashboard.
Most enterprises cannot, on demand, name the on-robot substrate, the foundation model, the training cloud, and the orchestration layer underneath each deployment. Without that map, capital allocation runs on the chassis price and the API rate. Substitutability cost is invisible until the renewal letter arrives.
The substrate qualification gate.
Carrier-grade qualification has a 40-year track record in the telecom industry. NTT Laboratories, BT Labs, and Deutsche Telekom Labs each run a multi-quarter cycle ending in a stamp that competitors spend two product cycles trying to dislodge. Marvell cleared NTT Labs on its broadband platform; the qualification carried across chassis and customer cycles. No equivalent gate exists today for the on-robot compute substrate or the inference ASIC layer.
The second-source mandate.
Carrier and OEM procurement standards require a named second-source path before a layer commits to a single-vendor supply. In 2026, the foundation model and on-robot substrate layers at most enterprises have no second source.
A chassis written off is a quarter of OPEX; a substrate written off is a board event. The instrument gap is what makes the gap material.
Pinpoint where growth hits a wall. 🚧
4️⃣ Market Reality: Evidence, Not Hype
Public, verifiable data, consistent with the May 12 and 19 The Pulse™ coverage and the May 14 and 21 The Vector™ deep dives.
Inference cloud commitment: Anthropic committed $200 billion over five years for 5 GW of Google Cloud TPU capacity, first capacity in 2027.
Inference revenue concentration: Anthropic and OpenAI contracts now account for more than half of the cloud providers’ approximately $2 trillion revenue backlog.
Custom inference silicon: Broadcom reported 45% growth in custom ASIC revenue, compared with 16% for the GPU segment in the same quarter.
Sovereign-edge federation: Vodafone, Deutsche Telekom, Orange, Telefónica, and TIM advanced a federated EU telco-edge cloud for sovereign AI workloads in mid-May 2026.
Physical-AI deployment: BMW Spartanburg ran a Figure 02 humanoid through production of more than 30,000 X3 vehicles at 99 percent placement accuracy per shift.
Physical-AI capital: Apptronik closed a $935 million Series A-X round at a $5 billion valuation in February 2026; Toyota TMMC signed Agility Robotics at roughly $30 per hour.
On-robot substrate lifecycle: NVIDIA Jetson Thor with IGX Thor has a 10-year lifecycle; Qualcomm Dragonwing IQ10 delivers 700 TOPS sparse, with 30% better power efficiency.
Standards lag: ISO 25785-1 for dynamically stable robots remains a working draft; ISO 10218-1:2025 covers collaborative applications but not walking robots.
Where data is not yet public, this briefing labels it as such. The hyperscaler internal-versus-external inference cost gap is the most cited unverified number; a published comparison showing that internal ASIC costs are more than 30% below external cloud rates would be the inflection point.
Credibility. Without this, no one pays. ✅
5️⃣ Strategic Implications: Four Lenses for Board Risk
a) Revenue Risk
Enterprises shipping AI features at scale face inference costs per token that grow nonlinearly above 500 million tokens per day on GPU-first architectures. Manufacturers shipping AI-assisted production face per-cell reintegration cost if the foundation model changes mid-cycle. The revenue at risk is not the line item; it is the gross margin envelope that the substrate decision locks for the next product cycle.
b) Cost Structure (load-bearing lens)
The procurement clock and the lock-in clock are mismatched in both domains. The GPU contract renews every 12 to 18 months; the silicon commitment behind it runs five to seven years. The chassis is replaceable every 12 to 24 months; the on-robot substrate lasts 60 to 120 months. The Wi-Next experience in Northern Italian factories is the operational lesson: the cell-level reintegration cost when an upstream layer changes is the line item most underestimated in a pilot business case, by a factor of two to three. Cost-per-token at production and cost-per-cell-hour in the plant must appear on the dashboard before procurement closes, not after.
c) Supply Chain Exposure
In inference, custom silicon at hyperscaler scale is concentrated among two merchant vendors and three closed-loop hyperscalers. In physical AI, Morgan Stanley flags a likely chassis shakeout, with embedded model and training-pipeline costs harder to unwind than the chassis itself. The substrate duopoly (NVIDIA and Qualcomm) limits negotiating leverage. Single-vendor exposure at the substrate or model layer is the supply-chain risk worth naming.
d) Competitive Positioning
OEMs that commit to a substrate first set the reference architecture. BMW’s Center of Competence for Physical AI in Production, established in February 2026, is the institutional signal that an OEM-level standard is being written this year. The AVCC governance-table convention, which assembled OEMs, Tier 1 suppliers, and compute-stack vendors into a founding group, was built to hold exactly this inflection. The OEM that wins the substrate qualification rents a decade of upstream optionality. Late entrants pay the reference-architecture tax.
Tie technical issue directly to enterprise value. 🔗
6️⃣ Scenario Analysis
Three scenarios for the 18-month horizon, sized across both inference and physical AI. Probabilities sum to 100. Each is anchored to one observable trigger.
Base Case is the modal outcome, but not a comfortable one; even at base, the cost of running blind through the substrate decision exceeds the cost of building the dashboard.
Give executives a map of possible futures. 🗺️
7️⃣ Strategic Options: What Can Actually Be Done
Four options for a board or executive team in the next 90 days. All four can run in parallel; the priority column names what to authorize first.
Substrate qualification and the second-source mandate compound across product cycles. Both replicate procurement disciplines telecom has used since the 1990s. The Marvell NTT Laboratories qualification was not a contract win; it was a credential that carried across chassis and customer cycles. The same discipline applied to the inference ASIC or the on-robot substrate builds a governance asset, not a contract.
Convert insight into actionable strategy. ⚙️
8️⃣ What Leading Players Are Doing
Seven moves are visible in the public record over the past 6 months. Each can be sourced to an earnings call, a press release, or a regulatory filing.
NVIDIA. The cross-domain position. NVLink Fusion absorbs third-party ASICs into the inference fabric; Jetson Thor with IGX Thor carries a 10-year lifecycle; Isaac GR00T contributes the foundation model as a public reference architecture. The OpenFog and IEEE 1934 pattern at commercial scale.
Broadcom. Only public merchant silicon vendor at hyperscaler ASIC scale; 45% custom AI revenue growth versus 16% for the GPU segment. The next earnings call is the trigger watch.
Qualcomm. Physical-AI substrate challenger with Dragonwing IQ10. An open VLA bid would split the foundation-model layer.
Google. TPU gen 3 internal inference cost reported below external cloud rates; Anthropic anchor at $200 billion over 5 years. The closed-loop ASIC proof point is now cited in GPU renegotiations.
BMW. Center of Competence for Physical AI in Production, February 2026. Spartanburg pilot delivered 30,000 X3 vehicles at 99 percent placement accuracy. Healthy stack profile: one substrate trend, two chassis vendors.
Mercedes-Benz. Funded Apptronik through Series A and A-X at a $5 billion valuation; Apollo is running in Germany and Hungary. Second OEM-as-standard-setter.
Toyota Motor Manufacturing Canada. First public RaaS contract on a vehicle assembly line: 7 Digit units at roughly $30 an hour. The hourly figure converted humanoids into a manufacturing OPEX line item.
Where verification is incomplete, the briefing labels the inference. Internal hyperscaler ASIC cost comparisons and specific OEM design-win commitments remain unconfirmed and are labeled as such above.
Others are already moving. 🏃
9️⃣ Board-Level Questions
Seven questions for the next board meeting. The goal is not to extract answers in a single session; the goal is to install the substrate decision as a standing line item on the audit and capital allocation agenda.
Across every active AI deployment, can the executive team name the substrate, foundation model, training cloud, and orchestration layer? When does that map land at the audit committee?
What is our combined substrate lock-in horizon if every procurement under negotiation closes this calendar year?
Where do we have a named second source at the substrate and foundation-model layers, and where do we not?
Which named individual on our team is qualified to negotiate at the substrate layer in inference, and the on-robot compute layer in physical AI?
What is our governance trigger for substrate-strategy reviews: annual, quarterly, or event-driven?
How much of our compute and physical-AI spend rolls into substrate commitments versus chassis, model, or cloud?
If a chassis vendor consolidates, a hyperscaler tightens pricing, or a safety event stops an OEM line in 2027, what is our lever, and how long does it take to pull it?
Help them run better board meetings immediately. 📋
🔟 The Bottom Line
Two domains, one decision, one window. Substrate selection in 2026 is the next capital-allocation question for every enterprise running AI at scale and every manufacturer running a humanoid pilot. The chassis and the GPU are replaceable; the substrate is not. Companies that treat this as procurement will lose to those who treat it as strategic governance and bring the substrate dashboard, the qualification gate, and the second-source mandate to the board this year.
The Centillium precedent is the shorthand. Once NTT Laboratories qualified the EPON ONU, competitors spent two product cycles trying to dislodge it. The same discipline, applied to the inference ASIC and the on-robot substrate in 2026, builds a governance asset that compounds across cycles.
Clarity and conviction. ⚡
The next Sextant continues the Hidden Constraints series with the third constraint: the talent and engineering capacity required to govern the substrate decisions named here. From issue three forward, the series sits behind the paid subscription.
📣 Call to Action
The Sextant™ is your bi-weekly board-level briefing on the hidden constraints shaping deep-tech infrastructure decisions. Every issue takes a position, names the smoking-gun facts, runs the scenarios, and gives a CEO the questions to bring into the next board meeting. The product is not the document; it is the trust and pattern recognition that compound across briefings tied to your decision cycles.
If today’s issue earned its keep, the next will earn it again.
Subscribe to receive every Sextant™ in the Hidden Constraints series,
plus full access to The Pulse™ and The Vector™ extension.
Compound across briefings. 🚀
❓ Question for You
Across your active deployments, which substrate decision is now within 12 months of irreversibility, and who on your team is qualified to negotiate at that layer when the renewal letter arrives?
If you cannot answer the second half, the substrate decision is being made for you by your vendor, not by your board.
Carry it into the next board meeting. 🪑
🗺️ The arc so far
The Sextant™ (May 6): “The Hidden Bottleneck in AI Compute”. Board-level frame; first named the capacity constraint that made the bifurcation visible. The strategic-question anchor for Weeks 19 through 21.
The Pulse™ (May 5, Week 19): “AI Capex Sets the Clock”. Quantified the hyperscaler commitment surface that now anchors the Anthropic / OpenAI half of the fork.
The Vector™ (May 7): “The AI-RAN Architectural Fork”. Deep dive on the carrier-grade compute choice that the federated EU telco-edge now operationalizes.
The Pulse™ (May 12, Week 20): “Physical AI goes operational”. Deployment-wave frame; established that the stack had crossed into production, teeing up the architectural question answered this week.
The Vector™ (May 14): “The Inference ASIC Fork”. Silicon-level preview of the architectural bifurcation this week names at the full-stack level.
📅 Board Advisory Session
I work with a small number of deep-tech executives on a fractional basis: as a CxO who has sat on standards boards (IEEE, ATIS, TIA, EIA), run cross-functional GTM in Germany and Japan, and led the technical and commercial integration that produces IEEE standards from industry consortia.
If your board or leadership team is navigating an AI, edge, or autonomous systems decision in the next two quarters, a focused advisory session may be the fastest way to pressure-test the options.
Thanks Armando, that line 'Inference substrate decisions taken in 2026 carry compute architectures through 2031 to 2033' landed. The Reality is this is an architectural lock, not a procurement checkbox, and it already shows up in TPUs, NVLink Fusion and on-robot stacks like Jetson Thor with IGX Thor. Boards need a substrate dashboard and a named second-source before procurement closes, otherwise capex becomes a sunk governance event.