Wednesday, June 24, 2026

OpenAI unveils its first custom chip, built by Broadcom

OpenAI has launched its first custom-built inference processor, named Jalapeño, developed in collaboration with Broadcom. This chip is tailored for the specific needs of OpenAI’s inference systems, with early testing indicating a significant improvement in performance-per-watt compared to current state-of-the-art alternatives. The partnership was first announced in October 2025, and the move is seen as a strategy to lessen OpenAI's dependence on Nvidia's GPUs. OpenAI president Greg Brockman highlighted the company's focus on optimizing its entire infrastructure, including chip architecture and deployment systems, to enhance the efficiency and affordability of its AI models.

Facebook rolls out an AI companion app for creators

Facebook has introduced a standalone AI companion app for creators, reimagining its Creator Studio tool to help users grow their audiences on the platform. Currently in testing with select creators, the app integrates Meta's AI creator assistant, which offers personalized recommendations based on content style, performance, and audience engagement. This initiative aims to keep creators active on Facebook amid competition from platforms like TikTok and YouTube, while also reducing their reliance on third-party tools for content ideation and performance analysis.

Daya raises $2.4 million from Aptos and others to power African cross-border payments

Daya, a stablecoin-native payments company founded by Nigerian entrepreneurs, has secured $2.4 million in pre-seed funding to create a financial operating layer for African businesses engaged in cross-border transactions. The funding round was led by Hivemind Capital, with participation from Lattice, Alliance, Globelink, and the Aptos Foundation. Daya's platform aims to streamline the fragmented payment processes currently faced by African businesses, allowing them to manage payments, currency conversion, and treasury operations through a single dashboard that enhances transaction speed and transparency.

NTSB launches probe into fatal Texas Tesla crash

The National Transportation Safety Board (NTSB) has initiated an investigation into a fatal crash involving a Tesla in Katy, Texas, where 76-year-old Martha Avila was killed after a driver, Michael Butler, crashed into her home. Butler allegedly claimed to be using Tesla's Autopilot feature prior to the incident, but Tesla has countered that data shows he pressed the accelerator to the floor, overriding the Full Self-Driving software and reaching a speed of 73 mph. The NTSB's inquiry will likely involve reviewing logs from the vehicle's onboard systems to clarify the circumstances of the crash.

Figma now has AI motion graphics and shader tools

At its annual Config conference, Figma announced significant updates aimed at enhancing design and coding capabilities. The new features include AI-generated motion graphics, allowing users to create animations by describing them to a chatbot interface, and shader tools that enable custom visual effects directly on the canvas. Additionally, Figma introduced coding layers, which allow developers to modify project code without leaving the design environment, facilitating a more integrated workflow for full-stack development.

Traders are getting worried the chip stock bull market is cracking. How to protect yourself

Traders are increasingly worried about the semiconductor stock market as signs of volatility emerge. On Tuesday, options traders purchased 74,468 put contracts on the iShares Semiconductor ETF (SOXX), marking 1.5 times the 20-day average volume. The South Korean KOSPI Index, a key indicator for the global hardware supply chain, has seen three drawdowns exceeding 10% this year, with one drop nearing 20%. Despite a remarkable 300% rally from 2025 lows, concerns about the sustainability of this growth are rising, reminiscent of past tech market downturns.

Accelerating Transformers Fine-Tuning with NVIDIA NeMo AutoModel

NVIDIA's NeMo AutoModel significantly enhances the fine-tuning process for Mixture-of-Experts (MoE) models, achieving a training throughput increase of 3.4-3.7 times compared to the previous Transformers v5 framework. This improvement is coupled with a reduction in GPU memory usage by 29-32%, allowing for more efficient training of large models. NeMo AutoModel builds on Transformers v5 by adding features like Expert Parallelism and DeepEP fused dispatch, while maintaining API compatibility with HuggingFace Transformers, requiring only a single import line for users.

Bessent sees GDP growth booming again this year. Kalshi traders see little chance of that

Treasury Secretary Scott Bessent expressed optimism about U.S. GDP growth reaching 3% this year, citing a strong underlying economy. In contrast, traders on the prediction market platform Kalshi assign only a 14.2% chance that GDP growth will fall between 2.6% and 3.0%, favoring a lower range of 2.1% to 2.5%. Bessent's confidence may be influenced by his belief in the new Federal Reserve chairman, Kevin Warsh, fulfilling inflation and growth mandates. The consumer price index recently increased by 0.5%, with annual inflation at 4.2%, the largest gain in three years.

Figma bets on human judgment at Config 2026 while the AI powering its canvas belongs to someone else

At Config 2026, Figma unveiled significant updates to its design platform, expanding its canvas to incorporate code, animations, and 3D effects, aiming to retain users amid rising competition from AI-driven tools like Anthropic's Claude Design. The new features, including Code Layers and collaborative tools for sharing AI prompts, are designed to enhance user creativity while addressing the challenges posed by high inference costs from external AI models, which have pressured Figma's profit margins. With over 95% of Fortune 500 companies relying on Figma, the company seeks to maintain its leadership in the design space.

South Korea’s biggest chipmaker SK Hynix plans to raise $29 billion via Nasdaq listing as soon as July 10

SK Hynix, South Korea's largest company by market value, is set to raise approximately $29 billion through a Nasdaq listing of American depositary receipts (ADRs), with plans to issue 17.79 million new shares valued at 45.45 trillion won. The listing, expected to commence on July 10, aims to broaden the company's investor base and enhance its valuation, particularly as it seeks to capitalize on the booming demand for AI-related memory chips. SK Hynix currently dominates the high bandwidth memory market, holding about 60% of the share, and has seen its stock price increase over 280% this year, elevating its market capitalization above $1 trillion.

JPMorgan raises S&P 500 target to 7,800, but warns of 'flash crash'

JPMorgan has raised its year-end S&P 500 target to 7,800, up from a previous forecast of 7,600, driven by expectations of an AI supercycle and positive earnings revisions. The firm notes that while the U.S. and Iran are nearing a peace deal, which contributes to a favorable outlook, the market faces potential volatility, including a risk of a flash crash due to extreme speculative trading in secondary AI stocks. Despite these concerns, JPMorgan maintains that the fundamental economic backdrop is strong, advising investors to view technical weaknesses as buying opportunities.

‘Major refund opportunity’: IRS may owe millions of Americans cash — but you must claim by July 10 to get your chance

A federal court ruling has opened the door for millions of Americans to claim tax refunds related to penalties and interest assessed during the COVID-19 period. The National Taxpayer Advocate, Erin Collins, highlighted this as a 'major refund opportunity,' but emphasized that eligible taxpayers must file their claims by July 10, 2026. The court found that the IRS improperly levied penalties from January 20, 2020, through July 10, 2023, suggesting that these fees should be refunded. Taxpayers need to file Form 843 to request refunds or abatements, which cannot be submitted electronically.

OpenAI reveals its first AI processor: Jalapeño

OpenAI has introduced its first custom chip, Jalapeño, developed in collaboration with Broadcom, specifically for AI inference tasks. This ASIC (Application-Specific Integrated Circuit) is engineered to enhance the processing capabilities of large language models, including ChatGPT. The announcement comes just nine months after OpenAI's partnership with Broadcom aimed at reducing dependence on Nvidia's GPUs, which are currently in limited supply. Early testing suggests that Jalapeño will deliver performance per watt that significantly outperforms existing state-of-the-art chips, matching the capabilities of Nvidia's Blackwell and Google's Tensor processing units.

The global chip rout, Cerebras earnings, the PGA Tour's next chapter and more in Morning Squawk

Chip stocks plummeted across global markets yesterday, with Micron Technology leading the decline after one of its largest one-day stock drops in years. The memory chipmaker is set to report earnings later today, contributing to investor anxiety. Other major players like Intel and Advanced Micro Devices also saw losses, dragging the tech-heavy Nasdaq Composite down by over 2%. In South Korea, the market benchmark faced one of its worst days on record due to its high exposure to memory stocks, although some megacap tech stocks like Microsoft and Amazon managed to buck the trend.

OpenAI's deployment chief on Codex growth, falling AI prices, and the ROI question

OpenAI's deployment chief, Arnaud Fournier, revealed that the company's subsidiary DeployCo is embedding AI models into corporate IT systems, significantly enhancing user engagement with Codex, which now boasts over four million weekly users. Germany has emerged as a key market, experiencing a 720% growth in Codex users since January 2026. Despite the rising costs of newer models like GPT-5.5, Fournier noted that the overall price of AI intelligence has decreased by 100x in the past 18 months, driven by efficiency gains across the technology stack.

Qualcomm Buys Buzzy Chip Startup Modular for Nearly $4 Billion

Qualcomm has announced its acquisition of Silicon Valley chip startup Modular for nearly $4 billion, expected to close in the second half of 2023. The deal involves Qualcomm issuing up to 19.2 million shares, based on its last closing price. Modular, founded in 2022 by Chris Lattner and Tim Davis, specializes in a chip software platform and a proprietary coding language that enables developers to create AI software compatible with various chips. This acquisition reflects Qualcomm's strategy to diversify beyond its mobile device chip business and enhance its presence in the data center market.

Pangram CEO says language models give themselves away by making the same arguments

Max Spero, CEO of AI text detection company Pangram, asserts that language models expose their limitations by generating repetitive arguments. In an interview with AI Policy Perspectives, Spero described Pangram's deep-learning classifier as a black box, lacking full interpretability regarding its predictions. The tool detects suspicious phrases and identifies structural patterns left by language models, which even Pangram struggles to fully comprehend. Spero noted that while language models may perform better than average humans in grammar and logic, they produce a narrower range of arguments, contrasting with the diverse perspectives offered by human reasoning.

What to do if your iPhone is stolen – more detailed advice from Apple

Apple has enhanced its support page with detailed instructions for users whose iPhones are stolen, emphasizing the importance of security measures. Notably, the company warns against displaying contact information on the lock screen, as this could facilitate scams by thieves. Users are encouraged to activate Lost Mode immediately, which enforces biometric authentication for sensitive actions and prevents unauthorized access. Additionally, Apple cautions against removing the device from the Find My list, as this action disables Activation Lock, making it easier for thieves to erase and resell the phone.

November launch set for space shuttle Endeavour's towering display

The California Science Center in Los Angeles is set to unveil a stunning display of the space shuttle Endeavour on November 13, 2026. This exhibit features the shuttle stacked vertically with its external tank and twin solid rocket boosters, reaching a height of 184 feet, making it the tallest indoor museum exhibit in the world. After four years of construction, the Samuel Oschin Air and Space Center will also showcase over 25 aircraft and spacecraft in its various galleries, providing an immersive experience for visitors.

iPhone Ultra 3D-printed hinge problems reportedly solved, ready for September launch

Apple's iPhone Ultra, its first foldable device, is on track for a September launch as reports indicate that earlier issues with its 3D-printed hinge module have been resolved. The device is now entering test production, a critical phase before mass production begins. Concerns about hinge durability and a seamless design had previously threatened to delay the launch, with some reports speculating a release as late as December. However, a recent supply-chain report confirms that the hinge problems, supplied by Taiwan's Shinjuxing and American Amphenol, have largely been addressed, allowing Apple to proceed with its plans.

Samsung opens ChatGPT Enterprise and Codex access after AI restrictions

Samsung Electronics has expanded employee access to ChatGPT Enterprise and Codex, allowing all staff in Korea and Device eXperience employees worldwide to utilize these AI tools for diverse tasks. This decision marks a significant shift from the company's previous three-year restriction, which was imposed due to data-security concerns regarding the use of generative AI tools. The new deployment aims to enhance productivity in areas such as software development, marketing, and product development, leveraging AI for tasks like information search, document drafting, and coding.

Anthropic drops ‘workplace AI agents’ directly inside Slack

Anthropic has introduced its Claude Tag feature into Slack, allowing users to integrate AI directly into shared channels by tagging @Claude. This beta version enables team members to assign tasks, review outputs, and maintain context within discussions, marking a shift from isolated chat interfaces. Following a $65 billion Series H funding round, Anthropic's valuation now stands at $965 billion, surpassing OpenAI's $852 billion. The integration aims to streamline workflows by reducing the need for data transfer between chats and separate applications, enhancing productivity in enterprise environments.

Riot now lets you enable its anti-cheat when you want to

Riot Games has introduced a new feature for its Vanguard anti-cheat software, allowing players to activate it on demand instead of keeping it running continuously. This change is available to players of League of Legends and Valorant who have compatible hardware and opt into specific Windows security features. The Vanguard Pre-Check feature will not launch automatically at system boot, thanks to improvements made in collaboration with Microsoft's Xbox OS Security Team, which enhance the Windows kernel's defenses against cheat software. Currently, 35% of players meet the necessary requirements to utilize this feature, which includes running Windows 11 25H2 or newer.

Introducing computer use in Gemini 3.5 Flash

Gemini 3.5 Flash has integrated computer use as a built-in tool, enhancing its performance for agentic tasks. Previously available only in the standalone Gemini 2.5 model, this integration allows developers to create agents capable of interacting across browser, mobile, and desktop environments. The new capabilities support long-horizon tasks such as continuous software testing and enterprise automation. Additionally, Gemini 3.5 Flash includes safeguards against prompt injection risks, enabling enterprises to require user confirmation for sensitive actions and halt tasks if indirect prompt injections are detected.