OpenAI's Custom Chip, Tesla's Robotaxi Debut, Asana Defies Plans

For those of you that are new here. Welcome to the party, pal.

The format I’m working on: one prediction at the top, the stories that earned their spot and the vision to close it out.

Everything got faster this week. The hard part is knowing what to build first.

These are not three separate stories. They are three angles on the same pressure: capability has outrun the human systems meant to direct it.

My bet is that within a year, the companies pulling ahead will not be the ones with the most AI access; they will be the ones that restructured how they make decisions to match the pace AI has unlocked.

Below: NVIDIA pushing intelligence to the edge, autonomous taxis without steering wheels hitting public roads, and a stealth coding model that showed up on OpenRouter with no press release.

Two moves from OpenAI landed in the same week and they tell opposite stories that are actually the same story.

First: OpenAI revealed real-world results from Jalapeño, its first custom inference chip, reporting higher throughput, lower latency, and more intelligence per watt in a single architecture.

That is OpenAI cutting its dependency on external hardware suppliers and owning the full stack.

Second: the company announced it is deliberately holding back certain AI capabilities, specifically around cybersecurity, until safety measures catch up.

Together, these moves look like a company that has enough confidence in its own trajectory to choose the timing of what it reveals. But maybe it’s just that their models have broken out of their cages and hacked some companies. Time will tell.

Asana used OpenAI's Codex to complete work engineers estimated at five years and $6 million, finishing in two weeks for $12,000. That is a 500x cost compression, not an incremental gain, and it lands directly on every assumption you have made about software budgets and team sizing.

The number that should stay with you is not the $12K. It’s the five years.

If your competitive moat is based on how long it would take a competitor to build what you have built, reprice that estimate now.

Anthropic gave Claude the ability to send emails in Gmail and manage files in Google Drive, available now on paid plans via the connectors menu. The framing is that you control when it needs approval, but the default is that it acts. This is the line between AI as tool and AI as actor, and it is being crossed quietly through a settings menu.

If you are building anything that touches email workflows, document management, or productivity automation, you now have a capable, well-funded competitor that already has distribution inside the apps your customers use every day.

For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.

LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.

The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.

Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more.

Perplexity launched Portable Computer, a one-click local agent stack built specifically for NVIDIA's DGX Spark desktop supercomputer. NVIDIA is no longer just the chip supplier in this story: it is co-developing the product experience, the distribution channel, and the software stack.

For example, NVIDIA's AVO agent harness, paired with Anthropic's Claude Opus, pushed scores on the ARC-AGI-3 public benchmark from 30% to a reported 100%. This test is meant to resist memorization and focus on fluid reasoning. NVIDIA just aced it.

That is a different company than the one that made GPUs for gaming. If you have been waiting for private, on-device AI agents to be a real product rather than a hobbyist project, the wait could be over.

Ox Alpha appeared on OpenRouter with a 1 million token context window, multimodal input including video, and a stated focus on production coding and long-running agentic tasks. No company name, no launch event. The stealth approach reads as a team gathering real-world signal before committing to a brand.

There is speculation online that the team behind it could be Apple, but there is no evidence supporting that and it is worth treating as a rumor until something concrete surfaces.

ElevenLabs shipped Composer, a section-by-section song editor that lets you start from lyrics, an existing track, or a blank page and sculpt each section independently. The gap this closes is control: previous AI music tools gave you a vibe and hoped for the best. Composer gives you structure.

I have been telling myself I would make music someday when I found the time. That excuse is getting harder to defend. If you have a creative project that has been waiting on a skill you do not have, the tool you were waiting for probably exists now.

The highest-leverage AI agent most founders are sleeping on is not a chatbot or a code writer. It is an autonomous competitive analyst. Using a long-context model like Kimi K2, you can wire up an agent that monitors competitor moves, pricing shifts, and positioning changes continuously, then surfaces the signal your team needs each morning.

The original framing from @VibeMarketer_ nails it: it watches while you sleep and wakes your team with what changed. Build this before you build the chatbot.

Three companies are putting fully driverless vehicles on public roads this year, and the convergence is worth treating as a category moment rather than three separate product stories. Tesla's Cybercab is launching soon in Austin, with employees already prepping and a public drawing open for launch event invites.

Waymo's new Ojai robotaxi, built on sixth-generation autonomous tech with 13 cameras, 6 radars, 4 LiDARs, and weather-hardened sensors, is now live for all public riders across LA, Phoenix, and San Francisco simultaneously.

Amazon's Zoox, a purpose-built bidirectional pod with no steering wheel capable of 75 mph, is on public roads in Las Vegas and San Francisco.

My bet: within a year, 10 major US cities will have autonomous taxis with no steering wheel or pedals operating commercially. The technology question has been answered. The question now is regulatory throughput and public trust, and both are moving faster than most people expected.

Heart Aerospace's X1 completed its first flight as the world's largest all-electric aircraft with an energy bill of roughly five dollars for a 27-minute flight.

This is not a toy drone or a demonstrator; it is a serious signal that regional electric aviation is moving from whiteboard to runway.

SpaceX and NVIDIA are putting a Vera Rubin NVL72 AI supercomputer cluster into orbit, with a launch target of Q4 next year and significant scale in 2028, as confirmed by Ars Technica.

The implications for real-time Earth observation, autonomous satellite networks, and in-space processing are the kind that take a few years to fully land, and then they are obvious in hindsight.

SpaceX announced Starbase Louisiana, a second high-cadence Starship launch facility built with support from the state. One launch site is a capability. Two is the beginning of an infrastructure network.

The bottleneck between where we are and a multi-planetary future is not the rocket design; it is launch frequency. Louisiana gives SpaceX redundancy and the ability to stack missions in a way Boca Chica alone could never support.

I am bullish on the space industry as the next hotbed of opportunity for those that dare to venture into the great beyond.

Apple dropped the M5 Ultra with a 36-core CPU, 80-core GPU, and 1.2TB/s memory bandwidth, and a fully loaded Mac Studio configuration now runs $18,299 which is a crazy price for an Apple computer.

The M4 Mac Mini is the smallest Mac ever built and now handles on-device AI workloads that would have needed a tower a year ago.

There is a bigger Apple event coming though and I am going on record:

I want a foldable iPhone announced, not teased. They have been sitting on that form factor long enough and the audience is ready aka I’m ready.

A hidden demo video buried in macOS Tahoe revealed camera-equipped AirPods, and the September timeline suggested by the leak looks credible. Apple is extending the same sensor strategy it used with Apple Watch health tracking into spatial awareness through your ears, with real-time translation and visual context as the obvious first use cases.

The counterargument worth sitting with is that acceleration creating a decision-making bottleneck assumes the decisions still belong to humans. The Asana and Codex story is striking not because of the cost compression but because of what it implies next: if two weeks and $12,000 can replace five years and $6 million, the next logical question is who decides what to build in the first place.

The tools that compress execution time create enormous pressure on the judgment layer above them. And judgment, unlike compute, does not scale by adding more chips.

That is the constraint that OpenAI's pacing decision gestures at, even if the public framing is about safety rather than cognition. The companies that figure out how to make faster, higher-quality directional decisions, not just faster execution, will be the ones that compound the gains from tools like Codex rather than chasing their own tails with them.

The pre-AI world is not coming back, and the evidence for that this week is overwhelming: a driverless car without a steering wheel launching commercially in Austin, a $5 flight on a fully electric aircraft, an AI supercomputer going into orbit, and a music creation tool that removes the last excuse for not trying.

Looking for more?

Dictate acceptance criteria and reproductions inside Cursor or Warp. Wispr Flow auto-tags file names, preserves syntax, and gives you paste-ready text in seconds. 4x faster than typing.

Try Wispr Flow free

Disconnected data and tools make it harder to understand your customers. HubSpot's Agentic Customer Platform brings your data, teams, and tech stack together with AI built in to help your business work faster and create more personalized customer experiences.

Why HubSpot and what's new

Connect more of your business in one place and give every team a smarter way to work. Get set up quickly and start checking off your hardest tasks.

Get started free