OpenAI finished pre-training Bel, a base model above 10 trillion parameters, in the week after Nvidia agreed to fund $105B of Ohio compute. Scale and its sponsor arrived together.
Filings and related reporting say Nvidia will guarantee up to $105 billion toward a new Ohio data center campus with up to 4.25 gigawatts of AI compute, plus an option on a further 3.75 gigawatts. The chip supplier is now the balance sheet behind one of its largest customers.
OpenAI published first inference results for Jalapeno, its in-house ASIC, claiming industry-leading speed and efficiency. Secondary reporting puts the chip at 1.5x to 1.9x more work per watt, landing the week after Nvidia signed the funding deal.
Ottawa announced retaliatory duties on roughly $20 billion of US goods, reaching 50% across about 700 product lines. Hardware-heavy supply chains rarely enjoy a widening trade fight.
Koda publishes a fact-checked AI intelligence briefing every morning: site, email, podcast, and video. Founding sponsors lock today's rates for 6 months as the audience compounds.
View the media kit →Unconfirmed reporting on August 25 says OpenAI has completed the pre-training run for a base model codenamed Bel, carrying more than 10 trillion total parameters and succeeding the earlier 'Doug' pretrain. Bel is reported as a base for models beyond the upcoming Astra and GPT-6 class, which 'Doug' is expected to underpin after further reinforcement learning. The...
Continue reading arrow_outwardUnconfirmed reporting says OpenAI has completed the pre-training run for Bel, a base model of more than 10 trillion total parameters that succeeds the earlier 'Doug' pretrain and is reported as a base for models beyond the coming Astra and GPT-6 class.
Treasury Secretary Scott Bessent launched "Operation Economic Outcast" on 24 August, extending secondary sanctions to aviation, digital assets, gold, shipping and technology and designating more than 60 entities, including four India-based companies.
Mood reads Greed, with Nvidia's $105 billion commitment to OpenAI's Ohio campus reinforcing the view that compute financing, not model quality, sets the sector's ceiling.
OpenAI is now both Nvidia's largest financed customer and its most direct challenger: Jalapeño claims 1.5x to 1.9x more work per kilowatt than Blackwell while Nvidia writes the check for the Ohio buildout.
Bel is reported to carry more than 10 trillion total parameters, with the earlier 'Doug' pretrain, not Bel, expected to base the Astra and GPT-6 class models. The 'AGI-threshold' framing around the run is rhetoric, not a benchmark; what is reported is that the pretrain finished and the successor line has a base.
OpenAI's first published Jalapeño inference results claim industry-leading speed and efficiency, with secondary reporting putting the ASIC at 1.5x to 1.9x more work per kilowatt than Nvidia Blackwell on serving workloads. That is an attack on the CUDA serving moat by the same company Nvidia is funding with $105 billion for up to 4.25 gigawatts in Ohio, with an option on 3.75 more.
GPT-5.6 arriving in Amazon's Kiro at improved price-performance, following Sol price cuts of 20 to 33 percent, shows OpenAI buying developer habit as inference costs fall rather than hoarding the model inside Codex. The new Admin plugin for ChatGPT Work and Codex points the same way: the remaining enterprise blockers are provisioning, permissions and audit, not capability.
Upload an MP4, get music cut to its exact length, skip the licence hunt entirely; the catch is everything the tool refuses to accept.
This is for people shipping short, low-stakes video where clearing a music licence costs more than the clip is worth: internal explainers, product demos, social cutdowns. At $6.99/month for the Pro tier it is cheap enough to be a line item you forget about, and the video-first workflow genuinely removes the browse-and-audition step that stock libraries force on you. Adopt it as a utility, not as a music production tool, and do not plan a longer edit around it.
“Uploading an MP4 was straightforward, and the generated tracks matched the video length exactly, which saved the reviewer from hunting for the right music.”Automateed review
Rowy is a low-code platform whose AI Bot feature ships templates for image and voice bots, so the first working version comes from configuration instead of a repo. That makes it a fit for the internal tool nobody wants to staff an engineer on: a triage bot, an intake form with voice, an image tagger. Build the throwaway version first and see whether the workflow survives a week of real use.
Try it arrow_outwardBotric AI is pitched as a conversational agent for customer service operations, which means the value is measured in tickets deflected, not in demo quality. Pilot it on one narrow category where the answers are already documented, such as billing or password resets, and compare handle time against your current baseline. If it cannot beat your help centre on that category, it will not beat your agents on the hard ones.
Try it arrow_outwardFinalle.ai combines new-media data sources with generative AI to aggregate and analyse financial information in one place. It is worth a scouting pass if your morning routine is five terminals and a newsletter folder, but treat outputs as leads to verify rather than positions to take. Test it against a week you already know well and see how much of its analysis you would have acted on.
Try it arrow_outwardMeta's AI tools are aimed at businesses and developers building responsive products on top of its models, and the desktop push means it belongs in your evaluation set alongside the usual assistants. Read the directory listing before the marketing site so you can see the pricing and the alternatives in one view. The question to answer is whether the model layer is easy to swap out later.
Try it arrow_outwardRe-price your coding stack today. GPT-5.6 is now inside Amazon's Kiro at improved price-performance, following Sol cuts of 20 to 33 percent. Run the same eval suite across both and move the workload that actually gets cheaper.
Lock down access before you scale seats. OpenAI's new Admin plugin for ChatGPT Work and Codex gives centralized user management. Define owners, groups and revocation paths now, while your deployment is still small enough to fix.
Wire one retrieval connector this week. Copy the Guidepoint pattern - an MCP connector exposing more than 100,000 vetted transcripts inside Gemini Enterprise - and expose one trusted internal corpus to your assistant instead of building another chatbot.
Bel, if the reports hold, shows the parameter ceiling keeps moving, though the Astra and GPT-6 class is expected to come from the earlier 'Doug' pretrain. What changed this week is who pays for the power: Nvidia's $105 billion turns a supplier into a sponsor, while OpenAI's own Jalapeno benchmarks quietly argue the other side. For builders, the leverage is downstream - cheaper inference in Kiro, admin controls for ChatGPT Work, connectors that put real data in front of a model. Buy distribution and habit while the giants are busy buying gigawatts.
AI analysis, world news, markets, and tools. One briefing, delivered free.