Chip Crisis Escalates: Apple Massively Increases Prices, Stock Plunges

Chip Crisis Fueled by AI Boom: Apple Massively Increases Prices

Apple raised prices for Macs and iPads on Thursday, citing the exploding costs of memory and storage chips, which have become more expensive due to the AI boom. The base model of the MacBook Pro now costs $1,999 instead of $1,699, the entry-level MacBook Neo is $699 ($100 more than at its March launch), and the iPad Pro jumps by $200 to $1,199. According to analyst estimates, the prices for memory chips have quadrupled in the past year because AI chip manufacturers like Nvidia and AMD are demanding the components for their data center products. Micron and others are focusing on the more expensive data center chips, earning more and supplying fewer consumer goods. “We have never seen a component price rise so much and so quickly,” said an Apple spokesperson. Tim Cook had already announced “significantly higher memory costs” in April; the stock fell by up to 6 percent on Thursday. Microsoft had already increased Surface prices in April. → www.nytimes.com

Synthszr Take: The AI bill is now arriving in the living room, not just on the hyperscalers' balance sheets. When the world's most expensive hardware company raises its prices by up to 25 percent and openly attributes it to the memory hunger of data centers, the compute boom has arrived in the real economy. Micron, with an 853 percent stock increase in twelve months, is the hidden champion of this phase, as the bottleneck is no longer just GPUs, but the mundane memory alongside them. Apple's 6 percent drop shows the other side: The concern that demand will slump is justified, but for a company whose devices were never cheap anyway, the effect will likely be smaller than expected. The exciting part will be seeing who can pass on the component costs and who absorbs them; this will separate the brands with pricing power from those without. Anyone planning supply chains now should factor in memory as a scarce resource and not bet on a quick easing. The AI supercycle has a price, and in the end, the end customer is the one who pays it.

China's AI Infrastructure Reaches the West

Two movements are running in parallel, and both aim for the same thing: making Chinese models available outside of China. Moonshot is bringing its Kimi API to the AWS Marketplace, which significantly simplifies procurement for companies because billing and EDP drawdown run through the existing AWS relationship. At the same time, @teortaxesTex reports that Huawei might demonstrate a system on the scale of a 950 SuperPOD. This would mean that large domestic NPU clusters are being produced in significant quantities. If the numbers are correct, this noticeably improves the economics and resilience of China's model-serving, as the dependence on NVIDIA hardware decreases. One lever reduces sales friction in the West, the other reduces hardware friction at home. → AINews

Synthszr Take: A moat that the West had grown accustomed to is being undermined here. For years, the assumption was that China's AI was stuck in two places: sales to Western companies and access to NVIDIA compute. Kimi on the AWS Marketplace clears away the first problem, and a proprietary Huawei SuperPOD on a 950 scale clears away the second. We wrote in February that what happens in China doesn't stay in China; this is the next piece of evidence. For European companies, the consequence is sobering: the model layer is becoming interchangeable. Qwen and DeepSeek have long stood alongside Claude and GPT in the same stack map. Anyone who cleanly separates their architecture—model as an adapter, orchestration as proprietary logic—can test a Chinese model tomorrow morning without making a strategic commitment. True sovereignty doesn't lie in the model; it lies in the data you don't give away.

Developer Jobs: AI as a Job Catalyst, Not a Killer?

While tech layoffs in May reached their highest monthly level in years, with AI cited as the most common reason, a new analysis by venture firm SignalFire paints a contrasting picture. Across more than 80 million companies, engineering was the most resilient job function in 2025: while total hiring by large tech companies fell by 25 percent compared to 2019, engineering roles only declined by 11 percent. At the twelve “Tech Majors” like Alphabet, Meta, Nvidia, Stripe, and Apple, engineers even accounted for 55 percent of all new hires in 2025, up from 46 percent in 2019. Early-stage startups hired 7 percent more engineers than in 2019. This contrasts with Dario Amodei's warning that AI could wipe out half of all entry-level office jobs, while Anthropic's own chief economist, Peter McCrory, has yet to see a measurable AI effect in the labor market. Nvidia CEO Jensen Huang puts it this way: since all his engineers started working with agentic AI, they are “busier than ever.” → Techpresso

Synthszr Take: This is the Jevons paradox in its purest form, and it doesn't surprise me one bit. When syntactic coding costs almost nothing (an LLM works for about one dollar in token costs per hour, where an offshore developer charges 30), the demand for those who actually understand the problem explodes. This is exactly the movement we observed in India, where the six largest IT houses cut their new hires by 72 percent, while onshore expertise close to the context became competitive again. The 55 percent figure from SignalFire tells the same story from the other direction: building and understanding merge into one mind, and the Product Engineer replaces the old trinity of PM, designer, and coder. Anyone laying off engineering teams now and justifying it with AI is confusing efficiency with demand. The work isn't disappearing; it's shifting upward, toward judgment. Companies hiring people with taste, business acumen, and user empathy today are buying an advantage that no agent can certify.

Like Anthropic, OpenAI Also Bets on Slack as an Orchestration Channel

Since June 22, 2026, ChatGPT Enterprise can write in Slack, not just read. With the new connector update for Enterprise and Edu workspaces, OpenAI's assistant can join channels, upload files, create reminders, and change user profiles. The difference lies in the permission architecture: the old connector ran on read-only OAuth scopes, while the new one requires separate write tokens like channels:join, files:write, and users.profile:write. Granting these permissions not only expands the agent's visibility but gives it a different class of authority within the communication infrastructure. For large Enterprise Grid customers, the rollout depends on approval from a Grid admin. The security calculus shifts because a successful prompt injection can now trigger actions, not just expose data. OpenAI itself openly states that prompt injection will likely never be fully solved and that agent mode increases the attack surface. → AI Secret

Synthszr Take: With write access, the agent becomes a self-directed actor on the edge of the organization, a colleague with no supervisor and no conscience. This is exactly what I described in Code Crash: anyone who lets agents into their organization is practicing Teal self-management, whether they want to or not. OpenAI's honest point that prompt injection will probably never completely disappear is actually a gift, as it forces the right response. That response isn't a ban, but governance, as has long been standard for sensitive pathways: write permissions only for clearly defined use cases, audit trails for every action, and human review at critical points like profiles and file uploads. The action control panel in the workspace settings is not a convenience but the only brake against a one-time OAuth grant that allows everything. The lesson from ShadowLeak (patched in December 2025) is not that agents are too dangerous, but that the guardrails must be in place before access is granted, not after. Those who set this up properly tomorrow morning will get a productive colleague on their team. Those who skip it will take on a risk that has write access.

OpenAI is Building Its Own Chips

OpenAI unveiled its first proprietary AI chip on Wednesday: Jalapeño, a so-called “Intelligence Processor,” developed in a nine-month collaboration with Broadcom. President Greg Brockman describes the chip as part of a “full-stack” infrastructure strategy designed to make compute cheaper, faster, and more reliable. The architecture reduces data movement and balances compute power, memory, and networking; initial tests show performance per watt is “substantially better” than the current state-of-the-art. Engineering samples are already running workloads, including GPT-5.3-Codex-Spark, and the chip is intended to be compatible with all major language models. OpenAI had planned its own chip factories in 2024 but scrapped those plans and instead announced the Broadcom partnership in October, with the goal of rolling out 10 gigawatts of its own AI accelerators. Analyst Jeremy Roberts of Info-Tech sees this as a step against Nvidia dependency and a possible entry into the data center hardware market. → The Deep View

Synthszr Take: In February, the message was still “Folks, it's going to get expensive, very expensive.” Jalapeño is the logical response: if you want to control inference costs, you have to go all the way down to the silicon. What's interesting is the compatibility with third-party models, because with this, OpenAI is positioning itself not just as a model lab but as a compute supplier for the entire industry, directly in Nvidia's territory. This fits the pattern we've been observing with OpenAI for months: Guaranteed Capacity, its own GPU networking standard, and now the chip. Every layer of the stack is being occupied, and in the end, the company that supposedly preaches the “democratization of intelligence” sits as a gatekeeper at every level. The honest question for anyone bringing AI into production is: how much lock-in are you buying into when the model, software, and hardware all come from one house? Vendor neutrality isn't a nostalgic principle; it's the only leverage you'll have for negotiation in two years.

Figma Focuses on Teamwork with New Tools

At its Config conference in San Francisco on Wednesday, Figma unveiled a whole range of new tools that frame AI less as a single-user accelerator and more as a team capability. With Code Layers, code becomes a design material: teams work with repositories, generate variants via AI, and synchronize changes between design and production code without leaving Figma. New Motion and Shader tools bring animations and 3D effects directly to the canvas, while Figma Weave bundles over 20 AI tools into reusable building blocks. In addition, there are agents that take over repetitive tasks. CEO Dylan Field was unusually clear: the creative breakthroughs that let a company win won't be coming from AI anytime soon. Head of Design Loredana Crisan doesn't ask if AI will replace designers, but what becomes possible when designers get new materials. According to Figma, its tools are used by 95 percent of Fortune 500 companies. → The Deep View

Synthszr Take: Figma is hitting on a point that most AI roadmaps miss. While half of Silicon Valley is selling agents as a replacement for human minds, Figma is building AI as a layer into the workflow where people continue to make the value decisions. This aligns with what I've been observing for months: AI generates functional interfaces in seconds, and for that very reason, design becomes more important, not obsolete. When everything is best practice, best practice becomes the new mediocrity, and the difference is made where someone consciously deviates from the expected. Code Layers is the real lever here, because the loops between the designer and the developer (which I've seen for years as a disguised conveyor belt) simply disappear. The designer orchestrates several agents and brings everything together in the end, instead of pixel-pushing every state herself. Anyone who integrates their tools this thoughtfully, instead of just dumping automation on top, is building an advantage that will show in their output tomorrow morning, not just after the next strategy offsite.

Genspark Design: Claude 4.7 Turns Everyone into a Designer

Genspark has officially rolled out “Genspark Design,” connecting idea, design, result, and code in a single workspace. At its core is Claude Opus 4.7 with its vision processing and reasoning capabilities, which not only creates pretty pictures but also interprets the user's intent. The previous build preview is now fully integrated into Genspark Design, including an AI Designer for UI, posters, video, and HTML animations. The crucial point: the drafts can be directly converted into runnable code (via Genspark Code), allowing for the creation of landing pages or interactive content without a separate development step. Genspark sells speed as its main feature—five eighty-percent drafts beat one perfect one. The provider is upfront about its weakness: without brand assets, references, and real copy, the result is mediocrity. For the launch, there is a 50% credit discount for individual and team users, and a four-week free trial for enterprise. → Trendium.ai

Synthszr Take: This is exactly what I meant by the version 1.0 principle. When AI can spit out production code as quickly as a demo used to be created, the reason to build mockups at all disappears. Genspark makes the leap from a Figma slide to a clickable HTML prototype in minutes, and that's the real leverage, not the fancy visuals. The line between conceptualizer and designer dissolves, something small teams and startups will feel immediately. The honest limitation is stated in the newsletter itself: if you don't provide brand knowledge, you get average results. This aligns with what we've been writing about Claude since May: the model is only as good as the context you feed it. The tool license will be a commodity in 24 months; the difference lies in whether you're already building yourself tomorrow morning or still waiting for the design agency.

Google's Gemini 3.5 Pro is Delayed

Google is delaying the launch of its next frontier model, Gemini 3.5 Pro, from June to July, as Business Insider has learned. At the I/O developer conference in May, Sundar Pichai teased the model and promised a launch “next month,” but now the company is taking more time to gather feedback from early testers and make adjustments. The model is already running for select users on Google's Antigravity platform and has appeared on the benchmark site LMArena. Gemini 3.5 Pro is expected to be better at long-horizon tasks and operating agents. Google has also incorporated findings from the recent Flash 3.5 model. The pressure is high: while Gemini 3 exceeded expectations, Anthropic and OpenAI are pulling ahead in coding, the first real enterprise use case for modern AI. → Business Insider

Synthszr Take: A four-week delay is not news that should make anyone in the AI business nervous. What's interesting is why it's being delayed: Google is collecting real-world use cases from early testers instead of just releasing the model on spec. This is exactly what we've been describing for years as the right approach: validate real value before scaling. However, Google is still lagging behind Anthropic and OpenAI in coding, and that's the first use case that's actually making money in enterprises. Whoever comes in second here loses developers and thus the ecosystem that feeds the next generation of models. In April, Google aggressively lowered prices for Gemini 3.1 Pro; now it's buying quality with time. Both are legitimate levers, but against the velocity of the competition, they only help if the model delivers on coding in July.

Meta and the US Government: No Agreement on AI Model Review

According to a report by The New York Times, Meta has not yet agreed to submit its AI models to the US government for review. White House officials are applying pressure, as OpenAI, Anthropic, Google, Microsoft, and xAI have already agreed to such an evaluation. This leaves Meta as the only major provider standing apart. Meta spokesperson Francis Brennan downplayed the situation: “As we work through the details, we hope to sign the agreement soon.” No specific details of the review or a timeline were mentioned. This development is part of the growing concern within the US government about the safety and control of American AI systems, which we previously reported on at the end of April and in mid-June. → AI Secret

Synthszr Take: Five of the six major providers have signed; Meta is hesitating. This is interesting because Mark Zuckerberg celebrates his models as open source, so he supposedly has nothing to hide. If the weights are freely available anyway, a government review should be a mere formality. Perhaps Meta's concern isn't about the model itself, but about who will define the guardrails in the future and what doors this opens for further regulations. Brennan's “hope to sign soon” sounds like the polite version of stalling until the terms soften. What's clear is that whoever is last at the table is either negotiating the best deal or will end up standing alone. Meta should decide before the other five have written the rules for everyone.

Agencies Bet on AI Agents, Brands Turn to Insourcing

In Cannes this week, agencies are outdoing each other with tests, case studies, and partnerships centered on AI agents. WPP is testing an agent for planning and buying premium video inventory, Dentsu has partnered with Newton Research and is piloting transactional buying in the US via Dentsu.connect. Indie agency Butler/Till is building Claude agents with DoubleVerify to check inventory quality and automatically steer campaigns away from poor ad placements. Dept, with its Agent Studio, is renting out its own engineering workflows and underlying technology to clients. At the same time, brands are catching up: Hyundai is developing its own bidding agents with AI provider Chalice and its in-house agency Canvas, using the containerized solution OpenXBuild. When buying on OpenX's SSP, the CPM for online video dropped by 67 percent, and the cost per “High Value Action” (such as a dealer visit) fell by 20 percent in the pilot. Hyundai is reinvesting the savings directly into its media budget. → MyClaw Newsletter

Synthszr Take: Agencies are selling their clients the technology that will make those clients obsolete. Hyundai shows the math: 67 percent lower CPM, 20 percent cheaper conversions, all through an in-house agency plus a tech partner. If agentic media buying works like this, then the license layer (Newton, Chalice, Claude via DoubleVerify) is the same commodity for everyone, available for anyone to buy. What remains is the delivery logic behind it: brand voice, buying nuances, and domain knowledge in a RAG database that the agent actually uses. This is exactly what Dept is selling with its Agent Studio, and it's the only part that can't be brought in-house overnight. Any agency that just resells tool access is selling its own expiration date at a discount. The only viable position is the pipeline logic that Hyundai can't build itself in six months, and agencies should align their investments with that starting this quarter, not with the next license slide.