The Pulse™ | Week 30: Owning the inference layer
Five signals from the fog. Coverage window: Mon, Jul 13 to 17, 2026.
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)
👋 Welcome back to The Pulse™
The Pulse™ is your weekly operating brief for execs, founders, and investors. What changed across Industrial IoT, Telecommunications, Edge Computing, Autonomous Systems, and Artificial Intelligence, why it matters, and what to do next, all in one place.
🩺 This Week’s Pulse™
Five moves this week share one instinct: stop renting the compute layer; own it.
Meta confirmed its Iris inference chip enters production in September, built to keep recommendation queries off NVIDIA’s margin.
Nokia did the reverse, embedding NVIDIA GPUs into the radio with a commercial AI-RAN launch.
Qualcomm walked away from small-cell hardware to concentrate on owned edge-AI silicon.
Accenture spent $4.175 billion to own the software that secures industrial operations.
US regulators moved to own the safety envelope for autonomous edge cases.
The prize is control of the inference layer, wherever the workload sits.
Share this: The 2026 fight is not who has the most compute. It is who owns the layer where inference actually runs: the chip, the radio, the factory, the vehicle.
🏭 INDUSTRIAL IOT: Accenture spends $4.175B to secure the factory
As industrial AI wires operational technology into enterprise IT, the attack surface widens faster than the models improve. Accenture agreed this week to acquire a majority stake in Dragos plus all of runZero and NetRise, a combined $4.175 billion, moving from OT security services into OT security software.
Dragos was valued at about $3.2 billion; the combined businesses carry roughly $208 million in recurring revenue.
The three targets cover asset discovery, threat intelligence, and firmware risk across grids, pipelines, and plants.
OT cybersecurity software is a $27 billion market in 2026, projected to reach $59 billion by 2031.
The next wave of IIoT differentiation will be measured in certifications and secured assets, not features.
📡 TELECOM: NVIDIA moves into the radio
Nokia launched its commercial AI-native RAN platform on July 15, built on NVIDIA’s Aerial software and CUDA GPUs, and SK Telecom brought an AI-RAN pilot live the next day. The radio is becoming a GPU workload. Carrier qualification has always decided which radios ship; I saw that firsthand taking Marvell’s broadband silicon through NTT Labs certification, and the same gatekeeping now applies to AI-RAN.
Nokia targets spectral efficiency gains above 100% by 2028, with T-Mobile field trials before year-end.
Commercial service is planned for 2027; Nokia is tying its 5G and 6G roadmap to NVIDIA.
Ericsson is taking the opposite bet, prioritizing independence from a single compute vendor.
Carriers now choose a compute partner when they choose a radio, and that decision will shape 6G economics.
⚡ EDGE COMPUTING: Qualcomm abandons the small cell for owned edge AI
Qualcomm stopped selling its FSM100 and FSM200 small-cell chips to new customers, freeing engineering for 6G RAN and, more tellingly, for a full-stack edge-AI platform anchored by its Edge Impulse and Arduino acquisitions. The edge-as-substrate thesis is the one I helped codify as IEEE 1934 through the OpenFog Consortium; this week it turned into a portfolio decision.
The Snapdragon X2 Elite Extreme carries an 80 TOPS NPU, the highest in the consumer PC segment.
Inference-optimized silicon is a market above $50 billion in 2026, up from roughly $20 billion in 2025.
On-device models under 4 billion parameters now hit production quality at 20 to 30 tokens per second.
The edge is consolidating around owned inference silicon; commodity connectivity hardware is being abandoned from the top down.
🚗 AUTONOMOUS SYSTEMS: Regulators move on the edge cases
NHTSA Administrator Jonathan Morrison told autonomous-vehicle operators that interference with police, fire, and EMS is unacceptable and demanded fixes by the end of the month. Days later, Zoox recalled 105 robotaxis after one drove into heavy smoke from a Las Vegas fire.
Zoox decided on the recall July 7 and notified NHTSA July 8, one day before the letter.
A DC Council robotaxi proposal drew a July 13 hearing; Uber and Waymo are pushing competing federal rules.
The bottleneck has moved from perception compute to governance of rare, high-consequence events.
Dense-city licensing will hinge on who instruments edge-case failures, a governance problem I watched take shape across eight years leading the AVCC.
🤖 ARTIFICIAL INTELLIGENCE: Meta builds to escape NVIDIA’s margin
Meta confirmed its in-house Iris accelerator enters mass production in September, designed with Broadcom on TSMC nodes and aimed at recommendation ranking, the workload behind the Facebook and Instagram feeds.
Iris is part of Meta’s MTIA program; initial silicon validation completed in about six weeks with no major anomalies.
Meta plans to double datacenter capacity from 7 gigawatts in 2026 to 14 gigawatts in 2027.
Every query served on custom silicon is a query that does not pay NVIDIA’s margin.
The inference-ASIC fork is now a balance-sheet decision at hyperscaler scale, not a research bet.
🔥 3 Non-Obvious Takeaways
1. The commodity compute layer is being abandoned from both ends
Qualcomm walked away from small-cell hardware the same week Meta walked away from renting GPUs for its core workload. One exits generic connectivity silicon; the other exits generic rented compute. The undifferentiated middle, rented and vendor-neutral, is the layer nobody wants to own, and margins there will compress accordingly.
2. Meta and Nokia made opposite bets on the same vendor, for the same reason
Meta is building around NVIDIA to protect its margin; Nokia is building NVIDIA into the radio to differentiate its own. Both moves are about capturing the economics of inference rather than surrendering them. The question is no longer whether to use NVIDIA, but where in your stack its compute earns you a defensible position.
3. This week’s most important AI story was a cybersecurity deal
Accenture’s $4.175 billion OT purchase says the gating item for industrial AI is no longer the model or the sensor; it is securing the converged OT and IT surface that AI creates. As factories, grids, and pipelines connect, the attack surface, not the algorithm, becomes the constraint that decides deployment pace.
🗺️ The arc so far
This week extends a thesis The Fog Signal™ has tracked since spring: control of a defensible layer beats raw performance.
The Sextant™ (May 6): “The Hidden Bottleneck in AI Compute.” Named compute capacity as the binding constraint on the AI buildout.
The Vector™ (May 14): “The Inference ASIC Fork.” Traced the silicon split between training and inference economics.
The Sextant™ (May 20): “The Hidden Lock-In Beneath Inference and Physical AI.” Showed how that split hardens into vendor lock-in.
The Pulse™ Week 27 (June 30): “Purpose-built takes the lead.” OpenAI’s Jalapeno ASIC confirmed the fork.
The Pulse™ Week 29 (July 14): “The map is the moat.” Advantage descending to scarce physical position: spectrum, power, silicon.
This issue, Week 30 (July 21): Meta’s Iris and Nokia’s AI-RAN turn the fork into a balance-sheet decision; ownership moves to the inference layer itself.
🧭 Where to start
If deciding where in your stack to own compute rather than rent it is a live question for your leadership team, the Strategic Review Session is the right place to work it through.
❓ Question for you
Which of your four critical verticals, IIoT, Telecom, Edge Computing, or Autonomous Systems, has the weakest instrumentation for detecting when a compute layer shifts from rented to owned?
If you are not tracking your vendor-margin exposure on core workloads, your efficiency per watt, and the security coverage of your converged OT and IT assets, you are making 2026 procurement calls without the leading indicators that separate early movers from laggards over the next two quarters.