Capital allocation has shifted from bits-based businesses to atoms-based businesses over the last few years. Investors have been directing capital to overcome the physical-world bottlenecks of inference. There's an emerging cohort of atoms-based companies in the fields of energy, manufacturing, space, robotics, and beyond. These hard asset businesses are increasingly built and valued on bundles of cash flows rather than revenue multiples. This requires an entirely different approach to capital formation across the company lifecycle.
AI enables us to do the things we previously procrastinated over.
Adding more features isn't a guaranteed solution for struggling startups.
Small VC funds face pressure to grow so they can keep investing in their companies and retain influence in later rounds.
Deciding that a colleague opposes your plan because they fear change does not tell you whether their objection is correct.
A founder has to protect their focus, attention, and execution on purpose, even when it's uncomfortable.
Angel investors prioritize founder stories and conviction over data, investing personal capital without external pressure.
Starting in 2027, Microsoft will deploy Maia at multi-gigawatt scale and make it the main inference system for its products. The next-generation Maia 300 uses compute chiplets on TSMC's N2P process and targets long-context, multi-step agent workloads. Digitimes reports that the supply chain is already preparing for a surge in server builds in 2027.
AMD's Helion agentic AI rack is designed for flexible AI infrastructure that caters to hyperscale deployments and enterprise needs. It supports a range of configurations, including CPU-only and GPU-integrated systems, aligning with diverse AI workloads. Helion, part of an upcoming open blueprint by AMD for OCP, is set for release in November.
Physical Intelligence opened a beta API for its π robotics foundation models: upload demonstrations, fine-tune on Pi-managed compute, and run the result on your own robot through a hosted inference gateway. Early partners spanned hardware from Actor's autonomous wheel loader, running a commercial deployment with no operator onboard, to Almond's dual-arm Axol, which learned to deburr its own spacers in four days. Since most ran cloud-hosted inference, network connectivity was a pain point for some.
This deep dive uses the Emmy compiler framework to follow PyTorch operations down to CUDA one lowering stage at a time: Tensor IR, Loop IR, Tile IR, Kernel IR, and finally emitted code. It uses RMSNorm and matrix multiplication as worked examples. It also covers operator fusion, schedule selection, and online attention rewrites, so readers can see what Triton-style stacks do under the hood.
Sponsored
Ads get ignored on social media but not in TLDR! Reach hardware engineers and makers where they actually pay attention. Learn more about sponsorship opportunities [8].
Apple Silicon Macs have powerful GPUs that sit mostly idle in normal use. Minions combines them across machines into a virtual GPU cloud for large ML inference jobs, with fault tolerance and flexible task scheduling. The pitch is that teams can use hardware they already own instead of waiting for expensive, centralized GPU capacity.
Anthropic's Frontier Red Team ran Gemma4-26B-A4B across 15 off-brand Android TV boxes with 4GB of RAM each, using pipeline parallelism at 0.39 tokens per second, to test whether small models could replicate across cheap IoT hardware. A 9B model got inference running on 8 of 9 hardware types tested, and a Kimwolf-sized botnet would total only about 40 H100s of compute. By their trend line, small models reach the needed capability in simplified tests by year-end.
Apple's Home Collection, built with LG and Schneider Electric, covers cameras, thermostats, sensors, a doorbell, and a deadbolt. LG's lock uses Ultra Wideband to unlock as you approach and works locally over Matter and Aliro, while its indoor cam physically tilts away when someone's home. Every device supports Thread, and cameras keep recording to built-in storage if Wi-Fi drops. Prices run $199.95 to $399.95, with availability starting in December.
An Abilene-scale build requires roughly $17B for chips and servers, $3B for power, and only $2.4B for construction. Transformers, generators, and cooling equipment face long lead times and rising prices.
Walmart has spent a decade trying to automate about 200 US warehouses.
Each core runs a 4-stage pipeline and replaces work done by general-purpose CPUs.
The $199 watch has a color e-paper display, runs open-source PebbleOS, and lasts about 8 days per charge.
Crescent Island, Intel's data center inference GPU, will be sold as the GPU D-Series.
At OCP, Meta showed IcePack, a 102.4T liquid-cooled network rack, along with its IceCube switch platform.
Sponsored
AI agents need real access to do useful work. But when they inherit human credentials or standing privileges, even one routine task can create outsized risk. Our guide offers a solution.
In SECURING THE AGENTIC ENTERPRISE [5], you'll learn:
* How to identify risks of inherited credentials, standing access, and unpredictable agent behavior * How intent-based, just-in-time, and just-enough access can limit an agent's blast radius * How 1Password Privileged Access extends access governance to AI agents and other non-human identities
GitLab fixed CVE-2026-97229, a CVSS 10.0 integer overflow vulnerability in regex validation for CI/CD configurations. In certain conditions, an unauthenticated user could execute arbitrary code on a GitLab server. The fixes are included in CE/EE versions 19.2.8, 19.3.5, and 19.4.2. The release also addresses high-severity XSS, authorization, and response-cache poisoning vulnerabilities.
TFG London, owner of Hobbs, Phase Eight, and Whistles, detected a cyberattack on October 11. Several IT systems remain offline, website functions are limited, and orders placed on or after Sunday may be delayed. TFG has not found evidence that payment data was exposed. It is still checking whether customer data was compromised. Attack method, ransomware involvement, and restoration timing remain undisclosed.
Wordfence discovered a new vulnerability in the Ghostscript interpreter and rendering engine which could allow attackers to achieve remote code execution (RCE) using specially crafted PDFs. For WordPress sites to be vulnerable, an attacker must be able to upload PDFs which are then processed by a vulnerable version of Ghostscript to generate a preview or thumbnail. Other applications that use Ghostscript to generate previews or thumbnails from PDFs may also be vulnerable.
Doyensec found a chain in Windows' SFAPE VBS enclave, used by Defender's Enhanced Phishing Protection and loaded by LSASS. A repeated call to Initialize wrote a fixed 32-bit value without validating the destination. The researchers corrupted the enclave-range check, then used a ClipboardInput race to gain a controlled memset, which led to VTL1 read and write primitives, replacement of the valid-target table, and arbitrary calls to existing VTL1 code through NtContinue. Microsoft fixed CVE-2026-83498 in September's Patch Tuesday and blocked older SFAPE images through the VBS Code Integrity policy.
When attackers obtain AWS credentials, they frequently test them for access to other services of interest like SES, SNS, and Bedrock. Datadog examined three credential harvesting platforms and found that they utilize the ListFoundationModels, InvokeModel, and Converse API to determine access. It also found a script that checked for promotional credits available in the account.
2026.10.14·securitylabs.datadoghq.com·6 min read·TLDR InfoSec The DOJ raided more than 40 Texas residences in late September and seized 187 laptops used by North Korean IT workers, while two US facilitators were arrested for helping funnel over $3 million from about 70 companies. A new FBI advisory said the scheme has adapted, with operatives subcontracting jobs to freelancers in South Asia and West Africa and recruiting financially stressed Americans, including active-duty military, to host laptops or sit interviews. Exposed workers have threatened violence or data leaks to extract payment, and some are now chasing identities that qualify for security-cleared roles.
Sponsored
Watch this on-demand webinar from AWS and SANS Institute to explore how APIs, cloud-native development, and agentic AI reshape security risks. Learn strategies to evolve threat models, strengthen governance, and adapt controls. Watch now and explore AWS Partner
Lastwall provides identity security for defense and government organizations. Its IDCommand Suite combines zero-trust architecture, PKI, passwordless authentication, and risk scoring from 200 contextual signals. Quantum Shield secures data in transit with NIST-approved post-quantum cryptography.
A comprehensive open-source security tools suite for OSINT reconnaissance, web security testing, API security assessment, and compliance checking with AI-powered analysis.
CodeGraph is a tool to build pre-indexed code knowledge graphs which can be fed to AI agents to speed up code exploration.
The UK's AI minister, Kanishka Narayan, is drafting plans for compulsory safety reporting and monitoring for frontier AI, and a potential requirement that labs grant pre-release access to models for review. Narayan compared this to the Industrial Revolution and said frontier labs shouldn't release tools they can't guarantee are safe. Narayan also confirmed that he reached an agreement with Anthropic to provide the AISI preview access to Mythos again.
Anthropic warns that small AI models could soon self-replicate across consumer IoT devices, creating distributed botnets capable of running AI workloads. Anthropic distributed Gemma4-26B-A4B onto 15 off-brand Android TV boxes but could not demonstrate self-replication. On the hardware in question, inference would be a slow process, so defenders may be able to respond in time to stop any abuse.
Hardware hacker Joe Grand found implants in a Trezor One and a Trezor Model T bought from unauthorized sellers, contradicting Trezor's public claim that none of its devices had been compromised, despite Grand reporting the findings in August. The attackers replaced each device's microcontroller with malicious firmware derived from open source that displayed one of a dozen pre-generated seed phrases while passing integrity checks against a copy of the legitimate firmware stored on an external microSD card. Unlike the Ledger implants, which exfiltrated real seeds over a cellular link, these relied on predictable seeds, and Trezor's move to cut authorized resellers suggested the company feared a similar shift into trusted supply channels.
Sponsored
Ads get ignored on social media but not in TLDR! Reach security engineers and CISOs where they actually pay attention. Learn more about sponsorship opportunities [20].
Hackers remotely altered Danfoss refrigeration controllers at hundreds of European supermarkets, forcing freezers into defrost for up to six hours.
An OpenAI engineer was arrested for allegedly hiding a toilet-paper-wrapped audio recorder in an executive conference room TV's USB port, after a complaint showed he had queried a productivity tool for leadership meeting locations and a chatbot account tied to his spouse had asked how to build and obfuscate the device.
A crypto miner on a compromised Northwestern University build server exposed Racket's Apple and Windows signing keys, prompting the project to rotate both, have GlobalSign revoke the Windows certificate, and urge macOS users on versions 8.12 through 9.3 to reinstall before Apple revokes its certificate at semester's end.
Sponsored
Most can't, which is why low trust is the real bottleneck on enterprise AI.
Airia's State of Enterprise AI Risk [5] report lays out the 5 sources of AI risk and how to mitigate them. Read it to see how:
* Ungoverned AI is the risk that compounds [6] every other one * Model risk does not end at deployment * Vendor-embedded agents can access enterprise systems without appearing in your inventory * Most organizations cannot say who approved an agent's decision, or who was responsible for the policy behind it
2026.10.14·airia.com·Sponsor·TLDR IT Microsoft has announced plans to expand its Azure Maia custom silicon system to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for its products. Maia 200 is currently in production for workloads such as Microsoft Copilot, while the upcoming Maia 300 architecture is designed to deliver significant efficiency and throughput gains for long-context and multi-step agentic AI tasks.
2026.10.14·blogs.microsoft.com·6 min read·TLDR IT Users are reporting IPv6 connectivity failures linked to Intel Connectivity Performance Suite, causing intermittent problems with services including OneDrive and OneNote. Removing or updating the problematic software has resolved reported issues, highlighting how vendor networking utilities can create difficult-to-diagnose enterprise connectivity problems.
2026.10.14·theregister.com·3 min read·TLDR IT AI-generated code is making traditional line-by-line human PR reviews less practical, shifting quality assurance toward automated checks, guardrails, and test-driven verification. Developers can maintain code quality without reading the implementation by using pre-code rules, multi-model review bots, and user acceptance testing.
2026.10.14·scienceisneato.substack.com·12 min read·TLDR IT Software-correctness tools reduce the space for bugs through three approaches: raising the floor with correct-by-construction features like type systems and memory safety, pushing the walls inward with verification tools like SMT solvers, and detecting remaining bugs through model checking and fuzzing. Tools that rely on separate models risk implementation drift, where the model no longer matches the actual code. Hydro addresses this limitation by running correctness checks directly against deployed code.
2026.10.14·jhellerstein.github.io·11 min read·TLDR IT IT leaders share the initiatives they are prioritizing in the final quarter, from delivering previously committed projects to demonstrating measurable business value. The recommendations emphasize finishing meaningful work, improving productivity, and establishing credibility for next year's investments.
2026.10.14·cio.com·9 min read·TLDR IT Sponsored
Alice tells an AI agent, “List my customers.” The agent calls the MCP tool: get_customer_list_tool. Alice isn't mentioned. (The MCP doesn't know who Alice is, anyway.) Teleport [13] explains how to carry identity through the agent – securely and temporarily using JWTs.