The Pulse™ | Week 22: Capacity gets an address
Five signals from the fog. Coverage window: Mon May 18 to May 22, 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™
Three moves this week point the same way. Blackstone committed $5 billion to a joint venture with Google that sells TPU capacity as a service, on a platform separate from Google Cloud. The FCC cleared Lynk Global and Anterix to fold satellite coverage into licensed private 900 MHz spectrum. Waymo paused robotaxi service in four cities because its cars could not read flooded streets. Compute, coverage, and autonomy are each being pulled out of the generic central provider and re-sited as owned, conditioned infrastructure. The question for 2026 is no longer how much capacity you buy. It is where it lands.
Share this: Capacity gets an address in 2026. AI compute, network coverage, and autonomous operation are all being unbundled from the central provider and re-sited as owned, location-specific infrastructure.
🏭 INDUSTRIAL IOT: Humanoid robots cross from pilot to line
The clearest industrial signal of May is not a new platform. It is humanoid robots moving from demonstration into scheduled production work. Robotera raised more than 200 million dollars in a round led by SF Group and began thousand-unit deliveries in the second quarter. BMW’s Figure 02 units at its Spartanburg plant have logged roughly 1,250 operational hours on 10-hour weekday shifts and handled more than 90,000 sheet-metal components on the X3 line.
Humanoid pure-plays raised 2.9 billion dollars across 16 disclosed deals in the year to April 2026.
The deployable unit is the staffed shift on a named line, not the robot itself.
Cost each humanoid against the shift it replaces, not against a generic automation budget.
Over the next two quarters, the operators who win will be the ones who can name the line, the task, and the hours, not the ones who simply buy fleets.
📡 TELECOMMUNICATIONS: The FCC moves coverage into private spectrum
On May 18 the FCC approved an experimental license letting Lynk Global integrate its satellite direct-to-device network with Anterix’s licensed 900 MHz private broadband, spectrum that utilities and other critical-infrastructure operators already own. It follows the FCC’s May 12 clearance of SpaceX’s $17 billion purchase of 65 MHz from EchoStar. The pattern is consistent. Coverage is being unbundled from the public terrestrial tower and re-bundled into private, licensed spectrum. In my carrier go-to-market years at Lantiq and Marvell, where qualification at NTT Labs was the gold standard, the carrier held the license; now the enterprise increasingly does.
Anterix 900 MHz already underpins private networks for utilities and other critical-infrastructure operators.
Resilience and control, not headline price, drive the private-spectrum decision.
Treat licensed spectrum and satellite fallback as a procurement line, not an IT afterthought.
Expect more enterprise spectrum and satellite deals over the next 90 days as operators harden coverage, they can govern themselves.
⚡ EDGE COMPUTING: Inference moves onto owned locations
Akamai reported early commercial traction this month for its Inference Cloud, the distributed service it runs across more than 4,400 edge locations on NVIDIA’s AI Grid reference design. A frontier-model provider committed $1.8 billion over seven years for Akamai cloud infrastructure. The reading matches telecom. Inference is moving off the centralized region and onto sited capacity close to where data is produced. That premise is the operating basis of edge computing as a discipline, and it is what the IEEE 1934 reference architecture, which came out of the OpenFog work I helped lead, was written to standardize.
The market for inference-optimized chips is set to pass 50 billion dollars in 2026.
Latency, data residency, and cost now favor the specific edge location over the cloud region.
Map which workloads must run on owned edge capacity before renewing central cloud commitments.
Edge inference moves from pilot to budget line over the next two quarters as enterprises price location into the architecture.
🚗 AUTONOMOUS SYSTEMS: Waymo meets the geography problem
Waymo paused robotaxi service in Atlanta, Austin, Dallas, and Houston on May 21 after severe weather; one vehicle drove into a flooded Atlanta street and sat stuck for about an hour. The same day, Waymo suspended freeway rides after its cars struggled with highway construction zones. This follows a recall a week earlier and a May 13 expansion to more than 1,400 square miles across 11 cities, with the NHTSA and the NTSB both running active investigations. Across eight years founding and leading the AVCC, I watched the industry learn that autonomy is bounded less by the model than by the specific road, weather, and work zone in front of it.
Waymo still targets up to one million paid rides per week by the end of 2026.
The binding constraint is the operating domain, the exact street and condition, not fleet size.
Judge AV partners on geo-fenced reliability data, not on total coverage area.
Expect slower, condition-qualified expansion over the next 90 days as regulators press on operating-domain limits.
🤖 ARTIFICIAL INTELLIGENCE: Blackstone gives TPUs a landlord
Blackstone committed $5 billion in equity to a joint venture with Google that will form a new US company selling Google’s TPUs as compute-as-a-service, on a platform deliberately separate from Google Cloud. Google supplies the chips, software, and services; Blackstone brings capital and real-estate development; the venture plans its first 500 megawatts of capacity online in 2027. The structure matters more than the figure. A hyperscaler’s silicon is being sold through an independent operator backed by a property developer, a model the market already calls neo cloud. The scarce input is no longer the chip. It is the permitted, powered site to put it on.
The venture targets 500 MW of capacity online in 2027, with Blackstone leading site development.
AI capacity is now a real estate and power business as much as a silicon business.
Track interconnect queues and site permitting wherever your AI roadmap depends on new capacity.
Expect more chip-and-property ventures over the next two quarters as compute supply reorganizes around land and power.
🔥 3 Non-Obvious Takeaways
1. Real estate is becoming the AI moat
Blackstone did not join the Google venture for the chips. It joined for land, power, and the ability to develop sites at scale. Akamai’s owned edge grid and the humanoid line deployments show the same instinct one and two layers down. The AI constraint has moved from the wafer to the permitted, powered, connected location, and capital is repricing to match.
2. Coverage and compute are now bought the same way
The FCC’s Lynk and Anterix license and Akamai’s owned-edge inference are the same decision made one layer apart. The enterprise that once bought generic carrier service and generic cloud is now buying licensed spectrum and sited inference. Control and resilience, not unit price, drive both. Telecom and Edge procurement are converging on a single posture.
3. Waymo’s flood pauses are a costing signal, not a footnote
Waymo’s stuck cars and the humanoid line deployments tell one story. The last variable in autonomy is the specific operating environment, the flooded street or the exact production task. The deployable unit is the site, not the fleet. Budgets built on fleet scale rather than qualified operating domains will misprice both robotaxis and robots.
🧭 Where to start
If where your capacity lands has become a live decision for your leadership team, whether that is a permitted data-center site, a block of licensed private spectrum, or an owned edge location, the harder question is no longer how much to buy but where to put it. Come chat with us. Our free Introductory Session is built for exactly that conversation.
❓ Question for you
Which of your four operational verticals, IIoT, Telecom, Edge Computing, or Autonomous Systems, has the weakest instrumentation for detecting where your capacity is actually landing, rather than how much of it you are buying?
If you are not tracking power-interconnect lead times, the cost of licensed or private spectrum against generic carrier service, and edge-location latency against your central-cloud regions, you are making 2026 procurement calls without the leading indicators that separate early movers from laggards over the next two quarters.
🗺️ The arc so far
The Sextant™ (2026-05-06): “The Hidden Bottleneck in AI Compute”. First framed the AI capacity constraint at board level.
The Pulse™ (2026-05-12, Week 20): “Physical AI Goes Operational”. Established the deployment wave now reaching specific sites and lines.
The Vector™ (2026-05-14): “The Inference ASIC Fork”. Silicon-level view of the inference bifurcation.
The Pulse™ (2026-05-19, Week 21): “The Inference Fork”. Named the hyperscaler-core versus sovereign-edge split this issue grounds in real estate and spectrum.
The Sextant™ (2026-05-20): “The Hidden Lock-In Beneath Inference and Physical AI”. Board-level lock-in frame that the siting question now sharpens.