The Manifest
Custom AI silicon took center stage as OpenAI unveiled its Broadcom-built Jalapeño inference chip and Qualcomm pushed into data centers, while logistics saw a 12.6 billion pound bid for warehouse giant Segro and ocean rates stayed stubbornly high.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01OpenAI unveils its first custom chip, 'Jalapeño,' built by BroadcomTechCrunch
- 02Google bakes computer control directly into Gemini 3.5 FlashThe Decoder
- 03Anthropic says Alibaba illicitly extracted Claude AI model capabilitiesReuters
- 04Prologis rebuffed in £12.6bn bid for UK warehouse giant SegroThe Loadstar
- 05AI was supposed to kill engineering jobs, but new data suggests they're the most resilientTechCrunch
Models & Releases
5 storiesOpenAI updates GPT-5.5 Instant to better read user intent
OpenAI tuned its most-used ChatGPT model for better intent recognition and multi-turn context. If your team leans on the default model for quick tasks, expect fewer misreads, but re-test any prompt chains built around the old behavior.
Snowflake CEO finds GLM-5.2 competitive with Opus 4.7 at a fraction of the cost
A Snowflake benchmark puts Zhipu's GLM-5.2 near Opus 4.7 on coding at a fifth of the cost per token, though it burns nearly twice the tokens. Worth pricing out for high-volume internal coding work before assuming the frontier labs are the only option.
Baidu releases Unlimited OCR, a 3B model that keeps the KV cache flat for long documents
Baidu open-sourced a 3B OCR model that holds memory flat across long documents in a single pass. For anyone parsing contracts, customs paperwork, or invoices at scale, flat memory means predictable cost on long files.
Qualcomm enters the data center market with its Dragonfly C1000 processor
Qualcomm is pushing into data center compute with its own processor. More inference silicon options means pricing pressure downstream, which matters if compute is a real line item in your AI roadmap.
Gradium launches real-time speech translation models beating the GPT realtime baseline
Gradium's models claim better accuracy and latency than the GPT realtime baseline across 20 language pairs. Relevant if you run multilingual support or supplier calls and have been waiting on usable live translation.
Supply Chain & Ops
5 storiesSamsara tracking label targets the $35B cargo theft problem
Samsara's new label adds real-time freight visibility across any carrier with no system overhaul, aimed at a $35B theft problem. Low-lift visibility upgrades like this are far easier to pilot than a full TMS change.
Ocean shipping recovery still a ways off despite US-Iran ceasefire
Spot ocean rates are expected to keep climbing for about four weeks with many vessels full into July. Budget for elevated rates through the summer peak rather than a quick reversion.
FedEx launches life sciences suite as healthcare push lifts earnings
FedEx is chasing high-value pharma and healthcare volume as general parcel softens. For temperature-sensitive shippers this likely means better-supported service tiers, worth a look at renewal.
Intermodal rail volumes rise as shippers switch from trucks
Shippers are shifting volume from trucks to intermodal in the latest weekly data. If truckload capacity tightens later this year, a pre-qualified intermodal lane gives you a release valve.
Gemini carriers sail past 90% schedule reliability targets in April and May
Maersk and Hapag-Lloyd's Gemini network hit its 90% on-time target across all trades. Reliability that high changes safety-stock math, so revisit buffers on those lanes.
Deals & Market
5 storiesAgility Robotics plans to go public via SPAC in a $2.5B deal
The humanoid robotics maker expects $620M in proceeds to fund deployment. Warehouse and manufacturing pilots are the proving ground, so watch where the capital actually lands before betting on the timeline.
Qualcomm to acquire AI startup Modular
Qualcomm is buying Modular to deepen the software stack around its chips. Consolidation like this narrows the independent tooling field, worth noting if you depend on Modular's stack.
XCures lands $46M Series B to clean up messy medical records with AI
Records normalization is an unglamorous but real use case for AI. The same pattern applies to messy supplier and parts master data, where cleanup quietly unlocks better planning.
Volkswagen sells majority stake in Everllence to Bain Capital
VW is shedding a non-core unit to streamline its portfolio. Large industrials continuing to divest is a theme to track for the stability of suppliers further down your chain.
Cerebras stock plunges after earnings as CEO says margin outlook was misunderstood
Shares fell on a narrower margin forecast in the chipmaker's first post-IPO report. A reminder that AI hardware economics are still being priced, which feeds into compute cost volatility.
Research & Frontier
4 storiesIBM claims world's first sub-1 nanometer chip technology
IBM's nanostack transistors aim to boost performance or energy efficiency. Years from production, but it sets the ceiling for where compute cost and density head next.
DFlash speculative decoding drafts whole token blocks for up to 15x higher throughput on Blackwell
UC San Diego's method drafts whole token blocks in parallel instead of one at a time. Faster decoding lowers inference cost, which eventually shows up as cheaper or quicker AI features.
RIFT-Bench: dynamic red-teaming for agentic AI systems
A new benchmark probes attack vectors specific to agents that take actions, not just chat. If you are deploying action-taking agents, this is the kind of evaluation to demand before trusting them with real workflows.
Critique of the Agent Model
A paper questions what actually counts as an agent amid heavy marketing of coding agents and co-scientists. Useful framing for cutting through vendor claims when you evaluate so-called agentic tools.
Org & AI Architecture
5 storiesCompanies scramble to stop employees from maxing out AI budgets on small tasks
The token-maxxing era gave way to token rationing as spend ran ahead of value. Set per-team usage visibility early, because unmanaged token spend scales faster than the payoff when no one is watching.
AI researchers continue to leave Google for its rivals
Top researchers Jonas Adler and Alexander Pritzel are moving to Anthropic, extending a talent drain. Where talent concentrates shifts which labs ship the next capability jump, worth tracking if you bet on one provider.
Big AI labs are hiring philosophers
Labs are bringing in philosophers to work on alignment and judgment questions. A sign that the hard problems are increasingly about values and reasoning, not just scale.
OpenAI's deployment chief on Codex growth, falling prices, and the ROI question
OpenAI wants to embed its own engineers deep inside large companies and points to fast Codex growth. The ROI question stays open, so anchor any rollout to a measured workflow rather than a headline number.
Meta employees warn AI moderation rollout is too fast
Staff caution that models now handle most moderation requests, with the share still climbing. A case study in pacing automation ahead of reliability, relevant to anyone automating judgment-heavy work.