the Now & the Next

A bi-weekly speculative fiction suggesting the shape of things to come.
(sourced from trustworthy trade pubs, think tanks + frontier science news)

115+ Signals Tracked
7 Collisions Detected
14 Industries
35+ Publications
2 weeks Signal Window

This fortnight brought 115+ signals across enterprise AI, and a clear pattern emerged: the infrastructure is diversifying, the power grid is straining, and every platform wants to be the one that orchestrates your agents. Here are seven collisions we detected in the noise.

01
The GPU Monopoly Cracks

When everyone decides one chip architecture isn't enough

+

Intel and SambaNova announced a three-tier inference architecture — GPUs for prefill, SambaNova's reconfigurable dataflow units for decode, and Xeon CPUs as host — explicitly designed for agentic workloads that strain monolithic GPU systems. The same week, Intel and Google expanded a multiyear collaboration to co-develop custom ASIC-based infrastructure processing units, acknowledging that CPUs have become active bottlenecks in AI pipelines. IBM and Arm took a different route, enabling Arm-native AI applications to run on IBM Z mainframes through shared software layers — no hardware change, no code rewrite. NVIDIA itself partnered with Marvell to pursue the AI control layer, and Applied Materials introduced deposition systems for angstrom-era logic chips, pushing materials science into the gap that lithography can no longer close alone. The GPU isn't being replaced. It's being surrounded.

⚡ The Now

The AI compute stack is diversifying from GPU-only to heterogeneous architectures. Intel and SambaNova split inference across three hardware types. IBM and Arm bridge mainframes to mobile ecosystems. Applied Materials is pushing chip fabrication to atomic precision. Industry analysts now identify CPUs — not GPUs — as the active bottleneck in complex agentic workloads. The monolith is giving way to a consortium of specialized silicon.

→ What's Next

Heterogeneous compute becomes the default enterprise architecture within 18 months. Companies that designed their AI stack around a single chip vendor will face a migration as painful as the cloud migration was a decade ago. Procurement will need to evaluate not just GPU allocation but the full compute mix — CPUs, ASICs, dataflow units, and accelerators — matched to workload type. The winners will be the companies that treat compute diversity as a design principle, not a compromise.

Data Center Knowledge
Three-tier inference architecture using GPUs, RDUs, and CPUs — explicitly targeting agentic AI workloads that strain monolithic GPU systems.
Data Center Knowledge
Multiyear deal to co-develop custom ASIC-based IPUs for Google Cloud. CPUs are the newly recognized bottleneck in AI pipelines.
Data Center Knowledge
Arm-native AI apps running on IBM Z mainframes through shared software layers. No hardware swap, no code rewrite.
Network World
NVIDIA partners with Marvell to build the orchestration layer that sits above the silicon. The control plane, not just the compute plane.
Applied Materials
Atomic-level precision deposition for 2nm Gate-All-Around transistors. When lithography hits its ceiling, materials engineering takes over.
02
Power Is the New Silicon

The AI bottleneck moves from chips to kilowatts

+

Chevron and Microsoft entered an exclusivity agreement for a $7 billion natural gas power plant in West Texas — 2,500 megawatts dedicated to AI data centers. Amazon raised its Mississippi data center commitment to $25 billion, converting a former manufacturing plant into AI infrastructure and promising 2,000 jobs. JLL reported that less than 10% of U.S. data center capacity is actually AI-ready, with existing facilities unable to handle the power density and liquid cooling that GPU-heavy workloads demand. In Pennsylvania, over $100 billion in private AI and energy investments triggered local resistance even as state leaders embraced the boom. NVIDIA's Rubin GPUs face delays tied not just to chip validation but to power consumption and cooling optimization. And OpenAI raised $122 billion — the largest tech funding round in history — with infrastructure buildout as the primary destination. The constraint isn't whether you can design the chip. It's whether you can power the building it sits in.

⚡ The Now

Power availability has overtaken chip supply as the primary constraint on AI infrastructure growth. Hyperscalers are shifting from traditional grid contracts to direct power ownership — Chevron building a dedicated plant for Microsoft, Amazon repurposing factories, states competing for data center campuses. Less than 10% of existing capacity can support true AI-dense loads. The infrastructure gap is physical, not digital.

→ What's Next

Energy strategy becomes AI strategy. Within 12 months, every major AI infrastructure decision will be a joint venture between a tech company and an energy company. Data center site selection will be driven by power availability first, network connectivity second. Companies without a power roadmap will find themselves locked out of the compute they need. The new currency of the AI race is measured in megawatts, not FLOPS.

Data Center Knowledge
Exclusivity agreement for a $7B natural gas plant in West Texas. 2,500 MW dedicated to AI data centers. Hyperscalers building their own power plants.
Chain Store Age
Repurposing manufacturing plants into AI data centers. 2,000 jobs. The factory of the future runs models, not machines.
Data Center Knowledge
Existing data centers can't handle AI power density and liquid cooling demands. Only 22.8% of AI initiatives meet ROI targets in production.
Network World
Delays driven by power consumption, cooling optimization, and HBM4 validation — not just chip design. The physics is catching up.
Data Center Knowledge
Largest tech funding round in history. Valued at $852B. Primary destination: compute capacity and data center expansion.
Route Fifty
Over $100B in private AI and energy investments in Pennsylvania alone. Local resistance meets political enthusiasm.
03
The Control Plane Wars

Every platform wants to be the one that orchestrates your agents

+

ServiceNow declared itself the control plane for agentic business, embedding AI, data connectivity, and governance into every product and launching a Context Engine to give agents enterprise-wide awareness. Salesforce turned Slack into the front end for enterprise AI — the place where agents live, not just where people chat. Microsoft expanded Copilot's agentic tools into government clouds, including GCC-High and DoD environments, with Agent Builder and Copilot Studio for low-code agent creation. Adobe is reportedly building a Model Context Protocol server for Marketo, letting marketers manage campaigns through AI prompts instead of manual workflows. Kyndryl launched a full agentic service management offering aligned with ISO 42001, complete with governance, security validation, and runtime guardrails. And Gartner gave the timeline: by 2028, over half of enterprises will stop paying for traditional AI assistants entirely, replaced by autonomous agents that execute, not advise. The question is no longer whether agents will run the enterprise. It's whose control plane they'll run on.

⚡ The Now

A platform war has erupted over who orchestrates enterprise AI agents. ServiceNow, Salesforce, Microsoft, Adobe, and Kyndryl all launched agent management and governance capabilities in the same two-week window. Each is positioning itself as the single pane of glass for deploying, monitoring, and governing autonomous AI workflows. The convergence is unmistakable — and the competitive stakes are existential.

→ What's Next

The control plane vendor becomes the most strategic technology partner in the enterprise. Whoever owns agent orchestration owns the decision layer — what agents can do, what data they access, what actions they're authorized to take. Within 18 months, CIOs will face a "control plane decision" as consequential as the cloud platform choice was a decade ago. Expect lock-in dynamics, interoperability battles, and a governance standards war fought through bodies like ISO and the Model Context Protocol consortium.

No Jitter
All products now embed AI, data connectivity, workflow execution, and governance. Context Engine gives agents enterprise-wide awareness.
MarTech
Slack becomes the interface where agents live and work. The chat platform becomes the agent operating system.
Nextgov
Agent Builder and Copilot Studio now in GCC-High and DoD clouds. Low-code agent creation for the most regulated environments on earth.
MarTech
MCP server for Marketo Engage lets marketers manage campaigns via AI prompts. The UI becomes the conversation.
Network World
Agentic service management aligned with ISO 42001. Governance, security validation, and runtime guardrails for enterprise AI agents.
No Jitter
By 2028, over half of enterprises stop paying for AI assistants. Autonomous agents that execute replace tools that advise.
04
Physical AI Hits the Factory Floor

The agent workforce picks up welding torches and borescopes

+

FANUC and NVIDIA partnered to integrate Jetson edge modules and Omniverse simulation into industrial robotics — training robots in virtual environments before deploying them on real production lines. HII (Huntington Ingalls Industries) and GrayMatter Robotics signed an MoU to bring physical AI into U.S. Navy shipbuilding: autonomous surface preparation, coating, and inspection on hulls that currently require manual sandblasting and grinding. Waygate Technologies and GE Aerospace deployed automated borescope inspection templates for GEnx engines, embedding AI-assisted visual guidance directly into maintenance workflows. Ford's Parasol team used parametric coding to compress week-long vehicle design projects into hours. Halliburton acquired Sekal to build real-time automated drilling control. And Siemens opened a $220 million rail manufacturing facility in North Carolina integrating AI, robotics, and augmented reality from day one. These aren't pilot programs. They're production deployments in shipyards, engine bays, oil fields, and assembly lines.

⚡ The Now

AI agents are entering manufacturing environments that were previously too dangerous, too precise, or too physical for automation. Shipbuilding. Jet engine inspection. Oil well drilling. Vehicle design. Rail manufacturing. Each deployment pairs AI perception (vision, simulation, sensor fusion) with physical actuation — not replacing skilled workers, but extending their reach into tasks that are tedious, hazardous, or time-constrained.

→ What's Next

The "digital factory" becomes the "autonomous factory" within three years. Simulation-first deployment (train in Omniverse, deploy on the floor) becomes standard practice. Industrial companies will maintain digital twins not as planning tools but as continuous training environments for their physical AI workforce. The labor conversation shifts from replacement to augmentation — and the companies that frame it correctly will attract both talent and contracts.

Assembly
Jetson edge modules and Omniverse simulation integrated into FANUC's robotics portfolio. Train virtually, deploy physically.
Naval Technology
Physical AI for autonomous surface prep, coating, and inspection on Navy vessels. Augmenting the shipyard workforce, not replacing it.
GE Aerospace
AI-guided borescope inspection templates for GEnx engines. Standardized, automated, embedded in the maintenance manual.
Ford Media Center
Parametric coding compresses hundreds of design iterations into the time previously needed for one.
Journal of Petroleum Technology
Real-time automated drilling control. The DrillTronics platform can reduce well delivery time by 25%.
05
The Sovereignty Stack

When owning your AI means owning your chips, your models, and your fabs

+

Microsoft launched three proprietary AI models — MAI-Transcribe, MAI-Voice, and MAI-Image — to reduce its dependence on OpenAI. Faster, cheaper, and built in-house. The same week, Intel announced it would buy back a 49% stake in its Irish fab from Apollo for $14.2 billion — the only European facility running high-volume EUV lithography, now back under full Intel control. Intel also joined Elon Musk's Terafab initiative, a vertically integrated chip manufacturing project aiming to produce one terawatt of AI compute annually, with ambitions extending to space-based data centers. NVIDIA acquired SchedMD, the company behind Slurm — the open-source workload scheduler used by 60% of the world's supercomputers — raising immediate questions about whether open-source neutrality can survive inside a hardware giant. And DeepSig, a small U.S. startup, landed at the center of OCUDU, a DoD-backed initiative to build an open-source RAN baseband stack challenging Ericsson and Nokia. From models to fabs to scheduling software to radio access networks, the message is the same: dependency is risk.

⚡ The Now

Technology sovereignty is no longer a government talking point — it's a corporate strategy. Microsoft builds its own models to hedge against its most important partner. Intel buys back its own factory. NVIDIA acquires the scheduler that runs most of the world's supercomputers. Each move reveals the same calculation: if your AI capability depends on someone else's infrastructure, you don't really have AI capability. You have a lease.

→ What's Next

The "build vs. buy" decision expands to the full AI stack. Within 18 months, enterprises will evaluate not just which models to use but whether they control the silicon, the runtime, the scheduling, and the training data. Sovereign AI funds — already emerging in the Middle East and Southeast Asia — will accelerate as nations realize that cloud subscriptions don't constitute technological independence. The companies that own their stack will set terms for everyone else.

Computerworld
Three proprietary models — transcription, voice, image — built in-house. Faster and cheaper than OpenAI equivalents.
Data Center Knowledge
$14.2B to regain full control of Europe's only high-volume EUV fab. Sovereignty isn't free, but dependency costs more.
Data Center Knowledge
Vertically integrated chip manufacturing at terawatt scale. Tesla, SpaceX, xAI, and now Intel. The AI compute race leaves Earth.
Computerworld
Slurm runs on 60% of the world's supercomputers. Now NVIDIA owns it. Open-source governance meets hardware economics.
Light Reading
DoD-backed open-source RAN stack. AI replaces traditional pilot signals. A startup at the center of telecom's sovereignty battle.
06
Agentic Marketing Rewrites the Account

The $700 million signal that advertising runs on agents now

+

Publicis took Microsoft's $700 million global media account from Dentsu — not through a pitch, but through a strategic partnership to embed agentic AI across marketing workflows. Microsoft will deploy 365 Copilot to 110,000 Publicis employees and use Azure as preferred cloud. The same fortnight, Publicis acquired 160over90, pouring $500 million into sports marketing powered by AI and Epsilon data. Adobe reportedly plans to announce an MCP server for Marketo at Summit 2026, letting marketers run campaigns through AI prompts instead of clicking through UIs. Consumer brands are retooling: E.l.f. Beauty is shifting from traditional SEO to agentic- and generative-optimized search. SharkNinja plans to launch 25-30 products annually using AI-enabled data integration. Albertsons tested ChatGPT-powered advertising. And SeatGeek became the first ticketing platform to offer primary and resale tickets through ChatGPT, with Deal Score and seat-view imagery embedded in conversation. The ad industry's new creative brief starts with an API call.

⚡ The Now

Marketing is being rebuilt around agent orchestration and AI-native discovery. The Publicis-Microsoft deal isn't a media buy — it's an infrastructure partnership. Brands are moving from optimizing for human search behavior to optimizing for AI bot discovery. The shift from SEO to GEO (generative engine optimization) is accelerating as consumer attention fragments across conversational AI surfaces.

→ What's Next

The media agency becomes an AI systems integrator. Within two years, the winning agencies won't be the ones with the best creative — they'll be the ones with the best agent infrastructure. Media planning becomes agent orchestration. Buying becomes algorithmic. Measurement becomes real-time attribution across AI surfaces. The $700 million account move is the opening shot in a structural transformation of the advertising industry.

The Drum
$700M account won without a pitch. Agentic AI embedded across marketing workflows. 110,000 employees get Copilot.
The Drum
$500M invested in sports marketing vertical. AI and Epsilon data drive measurable outcomes across media, sponsorships, and content.
Consumer Goods Technology
E.l.f. shifts from SEO to agentic-optimized search. SharkNinja targets 25-30 AI-driven product launches annually.
Music Business Worldwide
First ticketing platform with primary and resale tickets in ChatGPT. Deal Score and seat views embedded in conversation.
Grocery Dive
Grocery chain testing AI-powered advertising. The retail media network meets conversational commerce.
07
The Vertical AI Acquisition Spree

Every industry is buying its way into AI capability — fast

+

Cisco announced its acquisition of Galileo Technologies to build AI observability for multi-agent systems — trust and monitoring for the agents themselves. Motorola Solutions bought HyperYou, a startup building agentic AI for 911 call centers that handles two-thirds of non-emergency call volume. Halliburton acquired Sekal for autonomous drilling. Trimble signed to acquire Document Crunch, an AI contract analysis tool used on over 10,000 construction projects. Bureau Veritas bought Lotusworks to serve AI data centers. OneStream was taken private for $6.4 billion by Hg — an AI-enabled CFO platform, public for less than two years before private equity pulled it back. And Broadcom struck deals with both Google and Anthropic to supply custom TPUs and 3.5 gigawatts of compute capacity through 2031. The pattern: every industry vertical is acquiring AI capability through M&A rather than building it internally, and the deals are getting larger and faster.

⚡ The Now

AI acquisition has gone cross-industry and cross-vertical at unprecedented speed. Networking companies buying AI observability. Defense contractors buying 911 AI. Oil companies buying drilling automation. Construction firms buying contract AI. The build-or-buy calculation has tipped decisively toward buy — and the premium for proven AI capability is climbing with every deal.

→ What's Next

AI-native startups become the most valuable acquisition targets in every industry. The pipeline of acquirable AI capability will thin rapidly as the largest companies in each vertical absorb the best startups. Expect valuations for AI-native companies with production deployments to surge 3-5x within 12 months. Companies that haven't started their AI acquisition strategy are already late. The industrial-era conglomerate returns — not through horizontal integration, but through vertical AI capability stacking.

Network World
AI observability for multi-agent systems. Trust and monitoring for the agents, not just the network.
Government Technology
Agentic AI handles non-emergency 911 calls (two-thirds of volume) and provides real-time translation.
Construction Dive
AI contract analysis for construction. Scans for risk provisions, payment disputes, compliance issues across 10,000+ projects.
CFO Dive
AI-enabled CFO platform taken private after less than two years public. $6.4B. Private equity sees the upside that public markets didn't price.
Network World
Custom TPUs for Google. 3.5 gigawatts of compute for Anthropic. Supply agreements extending through 2031.

Frontier Science Feeding the Machine

The research signals underneath the enterprise stories. These breakthroughs — in physics-informed AI, living robots, agent governance, post-quantum cryptography, and AI cognition — are the tectonic plates on which the business collisions above are riding.

Physics-Informed AI
Large physics models predict aerodynamics, structural loads, and thermal behavior in seconds instead of hours. GM and Jaguar Land Rover are already using them. Simulation joins the AI stack.
Agent Trust
As agents move from chatbots to autonomous multi-step systems, the risks compound — misinterpretation, vulnerability to manipulation, unintended actions. Anthropic maps the trust architecture agents need.
Agent Governance
Real-time telemetry plus automated policy enforcement for multi-agent systems. Governance attributes embedded into the data stream, not bolted on after. The control plane gets its nervous system.
Distributed Training
Decentralized AI training across solar-powered nodes and idle GPUs. Federated learning meets renewable energy. What if the power constraint solves itself — from below?
Living Robotics
Tufts researchers created neurobots — tiny living robots from self-organizing cell clusters that wire their own neural circuits. Not machines mimicking life. Life, computing.
Post-Quantum Security
Formally verified ML-KEM implementation for post-quantum encryption. When quantum computers arrive, the cryptographic foundations need to already be in place. Amazon is laying them now.
Bioelectronic Medicine
A thumb-sized implant sustains genetically engineered drug-producing cells under the skin for over a month. Electrochemistry keeps them alive. The pharmacy becomes the patient.
AI Cognition
Claude Sonnet 4.5 develops internal representations corresponding to human-like emotion concepts — and they influence task selection and ethical reasoning. The model doesn't just predict tokens. It forms something resembling feelings.
Neuroscience
Connectome-seq tags individual neurons with RNA barcodes and maps synaptic connections by sequencing. Neural connectivity becomes a data problem. The brain, readable.
Clinical AI
AI analyzes CT scans by simultaneously focusing on fine detail and overall context — the way a human radiologist works. Over 96% accuracy. Early detection, machine-scale.