The Pulse™ | Week 31: The price of the layer
Five signals from the fog. Coverage window: Mon July 20 to 24, 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 signals this week, one shared move. The industry stopped competing on how much compute it can build and started competing on how much intelligence it can deliver per unit of cost, power, spectrum, and engineering time.
AMD priced its answer in tokens per dollar.
NVIDIA shrank a world model onto a single edge GPU.
AT&T told investors satellite extends the network rather than replacing it.
Tesla’s filing showed a human still holds the wheel remotely.
Siemens bought the AI it lacked time to build. The buildout race is becoming an efficiency race.
Share this: The AI buildout race is quietly turning into an efficiency race. The winners will be measured in tokens per dollar, watts per inference, and bits per hertz, not raw capacity.
🏭 INDUSTRIAL IOT: Siemens buys AI into the design loop
Siemens agreed on July 20 to acquire Precision Innovations, adding AI-driven, shift-left planning to its chip and system design tools. The target automates design exploration and optimization, the slow front end of building industrial silicon. It follows Schneider Electric’s $3.1 billion move for Cognite on June 30. Two automation majors, one pattern: buy the AI layer rather than grow it.
Siemens adds AI-powered SoC design exploration through the Precision Innovations deal, disclosed July 20.
The economics are cycle time: AI compresses the engineering hours between concept and working silicon.
Watch which automation incumbent moves next; the industrial-AI roll-up is accelerating.
Expect more 9- and 10-figure industrial-AI deals this fall as incumbents buy time they cannot build.
📡 TELECOM: AT&T frames satellite as extension, not replacement
On its July 22 second-quarter call, AT&T chief executive John Stankey reaffirmed the carrier’s partnership with AST SpaceMobile and dismissed the idea that direct-to-device satellite operators will take material share from terrestrial networks. His framing: space-based coverage is complementary, a way to extend the footprint without building towers. AST SpaceMobile told regulators on July 15 it is deferring commercial direct-to-device service to early 2027 after launch delays.
Stankey positioned satellite D2D as coverage extension, not a competitor to AT&T’s ground network.
The economics are reach per dollar: satellite fills gaps terrestrial buildout cannot justify.
AST’s slip to 2027 keeps terrestrial the primary layer through 2026.
Direct-to-device stays a coverage story, not a capacity story, until the constellations fly at scale.
⚡ EDGE COMPUTING: NVIDIA runs a world model on one GPU
NVIDIA’s Cosmos 3 Edge, a four-billion-parameter world model that reasons about a scene and generates robot actions on a single Jetson-class GPU, moved from its July 15 Tokyo unveiling to developer availability and a fast-growing industrial coalition this week. FANUC, Yaskawa, and Kawasaki are among more than 25 signing on. The point is architectural: inference for physical AI is moving on-device, off the cloud round-trip. The fog computing reference architecture I co-authored as IEEE 1934, from the OpenFog Consortium, argued a decade ago that latency-bound intelligence belongs near the sensor, not a distant data center.
Cosmos 3 Edge runs vision reasoning and action generation locally, no internet round-trip required.
The economics are watts per inference: on-device models cut both latency and cloud cost.
Robotics and industrial vendors are standardizing on the edge model layer now.
On-device physical AI is 2026’s fastest-moving edge segment: the cloud trains, the edge decides.
🚗 AUTONOMOUS SYSTEMS: A teleoperator crash exposes the driverless gap
Tesla filed four new robotaxi crash reports with federal regulators, disclosed July 20. One shows a remote operator drove a Model Y into a hidden tree stump while recovering it from a dead-end road in Houston. The revealing detail is a label: Remote (Commercial / Test), the first time Tesla has used it across 22 NHTSA reports. A human was in control, from a distance. As founding president of the Autonomous Vehicle Computing Consortium from 2017 to 2025, I watched the industry learn that the operator, human or machine, must be named in the safety case. Ambiguity about who was driving is the whole question.
Tesla’s filing marks its first Remote (Commercial / Test) operator label; a teleoperator caused the crash.
The so-what: driverless services still lean on remote humans, and disclosure lags the marketing.
NHTSA’s end-of-July deadline on first-responder interference raises the transparency bar further.
Expect operator-type disclosure and teleoperation limits to become the AV regulatory battleground into the fall.
🤖 ARTIFICIAL INTELLIGENCE: AMD attacks NVIDIA on tokens per dollar
AMD used its Advancing AI 2026 event in San Francisco, July 22 and 23, to launch Helios, its first full rack-scale system: 72 Instinct MI455X GPUs, 18 sixth-generation EPYC Venice CPUs, and ROCm software in one liquid-cooled rack rated up to 2.9 exaflops of FP4 inference. The pitch was not peak performance. It was up to 30 percent more tokens per dollar than the comparable NVIDIA rack, priced near $5.25 million. Anthropic signed on, with a planned AMD equity investment of up to $5 billion.
Helios pairs MI455X GPUs and 2nm Venice EPYC CPUs; AMD sells it on tokens per dollar, not raw flops.
The economics are the message: inference cost, not model size, now decides hyperscaler silicon choices.
Anthropic’s commitment gives AMD the reference buyer it needed against NVIDIA’s rack-scale lock.
The 2026 accelerator fight moves from performance benchmarks to cost-per-token; watch hyperscaler dual-sourcing accelerate.
🔥 3 Non-Obvious Takeaways
1. The buildout race is becoming an efficiency race
AMD priced Helios in tokens per dollar; NVIDIA shrank a world model onto a single edge GPU. Same bet, opposite ends of the stack: the number that matters is the cost of a delivered inference, not the size of the cluster. Capacity was the 2025 story; unit economics is 2026’s.
2. Driverless and AI-native both still carry an asterisk
Tesla’s first Remote (Commercial / Test) label and AT&T’s complementary satellite framing admit the same thing: the fully autonomous version is not shipping yet. Read the caveats, not the category names.
3. Industrial incumbents are buying the AI layer because building it is too slow
Siemens for Precision Innovations and Schneider for Cognite are one move: acquire the AI capability rather than grow it, because design and operations economics will not wait for internal R&D. The industrial-AI roll-up will define automation M&A through year-end.
🗺️ The arc so far
This issue extends a thesis The Fog Signal has tracked since spring: the AI stack’s center of gravity keeps descending, from capacity to ownership to, now, economics.
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 Pulse™ (July 21): “Owning the inference layer.” Players moved to own their compute rather than rent it.
This issue, Week 31 (July 28): efficiency becomes the axis; the layer is now priced in tokens, watts, and hertz.
🧭 Where to start
If the shift from raw capacity to unit economics, cost per token, watt, and hertz, is a live planning question for your leadership team, the Introductory Session is the right place to pressure-test where you sit.
❓ Question for you
Which of your four critical verticals, IIoT, Telecom, Edge Computing, or Autonomous Systems, has the weakest instrumentation for detecting the shift from compute capability to compute economics?
If you are not tracking cost per inference, watts per inference at the edge, and spectral or design-cycle efficiency by site or product, you are making 2026 procurement calls without the leading indicators that separate early movers from laggards over the next two quarters.