AI News

OpenAI Jalapeño vs. NVIDIA: 23 AI Updates That Matter

Direct Answer

OpenAI's Jalapeño chip is a major vertical-integration move, but it is not an exit from NVIDIA. The first-party results show meaningful inference gains on three large public models, while OpenAI explicitly says it will keep deploying NVIDIA and other accelerators. The strategic change is leverage: OpenAI now has a chip it can tune alongside models, compilers, networking, and serving software.

The week's broader pattern is similar. NVIDIA is moving up the stack toward the open-model community. Apple is making unusually large local models possible on one machine. Google, OpenAI, Anthropic, and Perplexity are turning agents into browser, desktop, memory, and event-driven products. Cheap intelligence is spreading, but the scarce layer is shifting toward distribution, data, permissions, and reliable deployment.

The practical takeaway: do not redesign a stack around one launch headline. Benchmark one real workflow, preserve provider portability, and tighten permissions before enabling remote control, memory, browser use, or scheduled actions.

Watch the Roundup

Credits: the roundup, demonstrations, and personal commentary come from Matt Wolfe's video. The ChatGPT Sites segment was sponsored by OpenAI. Major product and performance claims below were checked against first-party announcements and live product pages on 6 September 2026. Results, prices, rankings, and availability can change.

What Is Confirmed, Conditional, or Still Early?

StoryStatus on 6 Sep 2026What the evidence supportsWhat it does not prove
OpenAI JalapeñoConfirmed first-party resultsCustom inference silicon with measured latency and efficiency gainsA general 104x advantage or the end of OpenAI's NVIDIA use
NVIDIA and Hugging FaceAcquisition agreement announcedNVIDIA announced a $12.93B agreement on 3 SeptemberThat the transaction had closed when this article was reviewed
Open models lead token volumeLive gateway snapshotOpen models can represent a large share of tokens on Vercel AI GatewayTotal market share, request share, or superior economics for every task
Gemini Omni tops video rankingEarly community resultStrong initial preference on the Arena leaderboardA durable lead across production workloads
Skild S1 demonstrationCompany-reported demoPromising one-demonstration robot learning behaviorIndependent validation across robots, tasks, and environments

Jalapeño Changes OpenAI's Economics, Not Its Supplier List

Jalapeño is OpenAI's first custom inference chip. Inference is the work performed after a model has been trained: processing prompts, generating tokens, and serving users. That distinction matters because the chip is not presented as a general replacement for every training accelerator.

OpenAI tested Jalapeño on GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. Its published ranges are 1.5 to 1.9 times more peak throughput per watt and 1.7 to 3.6 times lower end-to-end latency. The much larger 104.3x figure belongs to a narrower comparison: more throughput at the previous time-between-tokens point for one DeepSeek R1 configuration.

Read benchmark headlines literally: “104x more throughput at the same inter-token latency on one operating point” is not “104x faster at everything.” The broad result is still important, but its scope should travel with the number.

The strategic advantage is co-design. OpenAI can tune the model architecture, numerical formats, compiler, runtime, network, rack design, and serving policy together. It also gains negotiating leverage and another supply path. Yet OpenAI's own announcement says it will continue deploying NVIDIA accelerators and other hardware for training and inference. The better interpretation is portfolio expansion, not divorce.

The NVIDIA-Hugging Face Story Changed After Publication

When Matt's video was recorded and published on 28 August, NVIDIA's reported $12.9 billion purchase of Hugging Face had not been publicly confirmed by either company. On 3 September, NVIDIA announced an agreement to acquire Hugging Face for exactly $12,930,300,000.

NVIDIA says Hugging Face will remain open, continue supporting the broader ecosystem, and will not require NVIDIA compute. Those are meaningful commitments, but the incentives still change. The most important model repository and collaboration layer now sits inside a full-stack chip and infrastructure company, subject to the acquisition closing.

For teams, this is not a reason to abandon Hugging Face. It is a reason to keep deployment artifacts portable: mirror critical model files, record exact revisions and licenses, maintain reproducible conversion scripts, and avoid making one hosted registry the only path to production.

Open-Model Token Share Needs a Denominator

The video cites a Vercel AI Gateway chart in which open models represented 62% of token volume while closed models represented 62% of requests. That is possible because a smaller number of requests can consume more tokens. By 6 September, the live leaderboard had already moved again, illustrating why the chart should be treated as a daily platform snapshot.

Token volume can reflect long contexts, verbose outputs, agent loops, or inefficient task completion. Request volume can reflect smaller interactions. Neither measure alone establishes revenue, user preference, benchmark quality, or total industry adoption.

MetricUseful forCommon mistakeBetter companion metric
Token shareInfrastructure demandCalling it user market shareCompleted tasks and cost per success
Request shareInteraction frequencyAssuming equal request complexityLatency and token distribution
Benchmark scoreScreening candidatesPredicting production quality directlyPrivate task-set evaluation
Model priceEstimating unit costIgnoring retries and tool callsEnd-to-end cost per accepted result

Apple Makes Large Local Inference More Plausible

Apple's M5 Ultra can be configured with up to 80 GPU cores, 512GB of unified memory, and 1.2TB/s memory bandwidth. Apple reports up to 4.5 times the peak GPU AI compute of M3 Ultra. M6 targets a different class of machine, with up to 32GB of unified memory and nearly 30% more peak GPU AI compute than M5.

The headline benefit is memory capacity. Models that would not fit on one conventional consumer GPU may fit in unified memory, reducing the complexity of multi-GPU partitioning. The tradeoff is that capacity does not guarantee fast generation, broad framework support, or good economics.

Buy local hardware when privacy, offline availability, predictable utilization, or data gravity pays for it. Rent accelerators when workloads are intermittent, rapidly changing, or need CUDA-specific tooling. For many teams, a hybrid router remains the rational choice.

GLM and Qwen Compete on the Efficiency Frontier

Z.ai's GLM-5.3-Flash is a 320-billion-parameter mixture-of-experts model with 18 billion active parameters. It is the release behind the earlier 0x Alpha preview and is positioned for agentic work with a lower serving footprint than a fully dense model of similar total size. The creator's visual benchmark in the video is useful as a qualitative test, not a universal ranking.

Alibaba's Qwen3.8-Flash release combines a long context window with multimodal inputs and compatibility layers for OpenAI- and Anthropic-style APIs. Open weights make local or self-managed deployment possible, but “open” still requires checking the exact license, model size, quantization, hardware fit, and serving stack.

Evaluation layerQuestion to askEvidence to retainFailure signal
Task qualityDoes it solve our actual work?Blind acceptance scoresBenchmark lead, poor production output
LatencyHow long until usable completion?P50 and P95 task timeFast tokens, slow tool loop
EconomicsWhat does one accepted result cost?Tokens, retries, tools, computeCheap call, expensive workflow
OperationsCan we deploy and recover it?Version, license, runbook, fallbackOne fragile provider path

The right response to a crowded release week is not to migrate five times. Build a fixed evaluation pack, run each candidate through the same inputs, and promote only the models that improve a measured outcome.

Google Shipped a Connected Release Cluster

The Google updates look separate, but together they show a product strategy: multimodal generation, transcription, travel actions, trusted source material, and remote agent control are becoming connected workflow components.

Gemini Omni 1.1 Flash

Gemini Omni 1.1 Flash adds 360p previews, output up to 4K, scene extension up to 40 seconds, start-and-end frame control, and short video references. It launched through AI Studio, the Gemini API, and Flow. Its early Arena position was strong, but the vote count was still small enough that builders should treat the ranking as a promising signal rather than a settled result.

Gemini 3.5 Transcribe

Gemini 3.5 Transcribe supports streaming and prerecorded audio, speaker and word timestamps, more than 85 languages, and sub-second streaming. Google's announcement cites third-party word-error-rate measurements of 4.0% for streaming and 2.6% for non-streaming. Those averages do not replace tests on your accents, jargon, background noise, and speaker overlap.

AI Mode travel and Gemini Notebook

Google expanded travel workflows with price tracking across more than 180 countries, points-and-miles support, and U.S. English hotel-booking partners. Gemini Notebook can now use select Google Play Books that the user owns as expert sources. Collaborators need to own the same book, which preserves licensing boundaries while making purchased material more useful inside research workflows.

ChatGPT Sites, clearly labeled

The sponsored segment demonstrates ChatGPT Sites turning a spreadsheet into a hosted team dashboard with authentication, data storage, and web-connected research. The useful lesson is rapid internal-tool prototyping. The responsible next step is still conventional product review: permissions, data retention, auth roles, auditability, backups, accessibility, and an export path.

Agent Convenience Expands the Permission Surface

The most consequential product updates are not model scores. They are the interfaces that let models act while the user is away: remote control, event triggers, browsers, shared memory, and local-first agent computers.

CapabilityWhat changedOperational valueRequired control
Antigravity Remote ControlResume local agent sessions from a browserContinue work away from the machineProtect the linked local files, tools, credentials, and environment
ChatGPT Work eventsTasks can start from supported service eventsLess polling and manual handoffNarrow triggers, connector scopes, and action approvals
Claude Cowork browserA separate built-in browser can perform web tasksLonger browser workflows with isolated loginsSite-by-site access and prompt-injection awareness
Claude in ChromeBrowser agent reached general availabilityWork directly in existing web contextUse a dedicated profile for higher-risk work
Claude memoryOne user-controlled memory across chat and CoworkLess repeated contextReview, edit, delete, and limit sensitive memory
Perplexity Portable ComputerLocal-first agent orchestration on supported hardwareKeep more execution and data on-deviceExplicit cloud-escalation and connector policies

OpenAI also announced a temporary GPT-5.6 Sol API price reduction of more than 20%, bringing the published promotional rates to $4 per million input tokens and $20 per million output tokens. Subscription usage did not change. This is another reminder to separate API economics from consumer-plan limits.

Browser agents are exposed to untrusted instructions. Anthropic explicitly warns that prompt-injection risk cannot be eliminated. Keep payments, destructive changes, account recovery, publishing, and irreversible messages behind human confirmation.

The Creative Economy Is Negotiating With AI, Not Resolving It

Dr. Dre's comments about using AI as a studio tool are a cultural signal, not a policy settlement. A prominent artist can find value in ideation or production assistance while still leaving authorship, consent, compensation, and training-data questions open.

Stability AI's $76 million Series B, part of $232 million in total announced funding, included backing from Electronic Arts, Sony Music, Universal Music Group, and Warner Music Group. That investment suggests large rights holders want influence and economic participation in generative media. It does not mean the industry has accepted every model, dataset, or output practice.

TIME's annual AI list is an editorial ranking of influence. It will naturally differ from a technical leaderboard or a list of people who shipped the most impressive models. Treat it as evidence about institutional visibility and cultural reach, not a scientific measurement of contribution.

Skild S1 Shows Why Robotics Claims Need Harder Tests

The Skild S1 demonstration shows a robot learning a roughly ten-minute task from one video demonstration without task-specific fine-tuning. If the behavior generalizes, it points toward a powerful interface: demonstrate the work instead of writing a control policy.

The current evidence is still a company demonstration. Before treating it as a deployable capability, ask whether it survives different lighting, object positions, hardware, interruptions, mistakes, and safety constraints. Robotics performance should be measured across repeat runs with human interventions and failure severity recorded.

Robot testMinimum evidenceWhy it matters
RepeatabilitySuccess rate across many trialsA polished demo may hide variance
GeneralizationNew objects, positions, and environmentsMemorization is not adaptable work
RecoveryResponse to slips, occlusion, and interruptionReal environments are not scripted
SafetyBounded force, stop behavior, and human overridePhysical errors have physical consequences

A Seven-Day Adoption Plan

  1. Day 1: write one workload card. Define the input, accepted output, latency target, privacy level, current cost, and failure consequence for one repeated task.
  2. Day 2: build a 20-case evaluation set. Use representative work, difficult edge cases, and examples that previously failed.
  3. Day 3: compare three model paths. Test one premium API, one cheaper or open model, and one fallback. Score complete tasks, not attractive samples.
  4. Day 4: calculate end-to-end economics. Include retries, tool calls, human review, hosting, and failed outputs rather than comparing token prices alone.
  5. Day 5: map permissions. List every file, browser profile, mailbox, calendar, credential, and publishing surface the workflow can touch.
  6. Day 6: add an approval and recovery path. Require review for consequential actions, retain an audit trail, and confirm you can revoke access or switch providers.
  7. Day 7: run one bounded pilot. Measure time saved, accepted-result cost, interventions, and failure severity. Expand only after the evidence improves.

This process makes the week's launches comparable. A faster chip, cheaper model, new browser, or memory layer is valuable only when it improves a real task without creating disproportionate operating risk.

Video Chapters

TimeTopicTimeTopic
00:00Intro17:14ChatGPT Work sign-in
00:16OpenAI Jalapeño18:03ChatGPT scheduled tasks update
01:46NVIDIA buying Hugging Face18:31GPT-5.6 Sol pricing change
03:56Open-weight takeover18:42Claude Cowork browser
05:21Apple M5 Ultra19:18Claude in Chrome
07:01GLM-5.3-Flash and 0x Alpha19:39Claude memory update
09:13Qwen3.8-Flash20:21Perplexity Portable Computer
10:08ChatGPT Sites21:40Dr. Dre likes AI
11:42Gemini Omni 1.1 Flash23:11Stability AI investments
15:17Gemini 3.5 Transcribe23:44TIME100 AI influence
15:42AI Search travel updates25:56Skild S1
15:56Gemini Notebook update26:44Personal achievement
16:20Antigravity Remote Control28:33Thank you, with cameos

Verdict

The most important story is not that OpenAI “moved against” NVIDIA. It is that the AI stack is becoming vertically integrated and strategically redundant at the same time. OpenAI wants its own inference silicon and continued access to NVIDIA. NVIDIA wants hardware leadership and ownership of a central open-model platform. Apple wants large local workloads on its machines. Model vendors want their intelligence embedded in every work surface.

For builders, abundance changes the optimization target. The best system is rarely the one attached to the week's highest benchmark. It is the one that produces an accepted result at a predictable cost, retains a fallback, respects data boundaries, and can be operated by the team after the demo ends.

Track capability releases, but invest in evaluation, portability, and control. Those assets survive the next model launch, the next pricing change, and the next acquisition.

Sources and Links

This article preserves the video's official YouTube publication date of 28 August 2026 and records material developments through 6 September 2026, including NVIDIA's later acquisition announcement. Recheck live pricing, rankings, availability, and transaction status before making a purchase or deployment decision.

Common questions

Is OpenAI Jalapeño 104 times faster than NVIDIA hardware?
Not as a general claim. OpenAI reported 104.3 times more throughput at the previous time-between-tokens point for one DeepSeek R1 operating point. Across three public models, the broader reported ranges were 1.5 to 1.9 times more peak throughput per watt and 1.7 to 3.6 times lower end-to-end latency.
Did NVIDIA buy Hugging Face?
NVIDIA announced on 3 September 2026 that it had agreed to acquire Hugging Face for $12.93 billion. When Matt Wolfe recorded and published this video, the transaction was still an unconfirmed report. An agreement to acquire is also not the same as a completed closing.
Do Vercel AI Gateway numbers prove open models have won?
No. They are a useful daily snapshot of traffic routed through one gateway. Token share is not request share, revenue share, or total market adoption, and open models may consume a different number of tokens per request.
Is Apple M5 Ultra useful for running local AI?
It is a serious local inference option for workloads that benefit from very large unified memory, but price, software support, quantization, and sustained throughput still matter. Compare the complete workload and cost against rented accelerators before buying hardware.
Which model release should a builder test first?
Test against your own task set rather than a general leaderboard. Qwen3.8-Flash is relevant for long-context and multimodal workflows, GLM-5.3-Flash for efficient agentic work, and Gemini Omni for multimodal media generation. Measure quality, latency, retries, and total task cost.
Which updates create the biggest new security risk?
Remote agent control, browser automation, persistent memory, and event-triggered tasks expand the permission surface. Use separate profiles, narrow connectors, approval gates for consequential actions, audit logs, spending limits, and a reliable way to revoke access.
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call