AI Stock Market Turmoil: DeepSeek Raises $16 Billion as OpenAI Stumbles

Market Tremor: AI Investments Steady, Stocks Wobble

Amazon, Alphabet, Microsoft, Meta, and SpaceX with xAI plan to invest up to $828 billion combined in data centers, chips, and other AI infrastructure in 2026. Amazon is planning $220 billion, Alphabet up to $205 billion. Microsoft is looking at around $170 billion, Meta up to $145 billion, and SpaceX up to $88 billion. Several banks, including Evercore and Bank of America, expect AI investments to reach one trillion dollars in 2027.

On Friday, the Nasdaq fell 2.9 percent and the S&P 500 1.8 percent. For both indices, it was the weakest day since June. The stocks in the Nasdaq 100 together lost more than a trillion dollars in market value. A weak quarterly forecast from Intel contributed to the decline. Among chip stocks, Nvidia dropped 4 percent, Broadcom more than 5 percent, and Qualcomm 11 percent. Alphabet had already tumbled 12 percent on Thursday—the worst day in company history—wiping out around $500 billion in market value.

In the third quarter, four of the five tech giants reported negative free cash flow. Alphabet was $9.1 billion in the red—the second consecutive quarter—with quarterly capital expenditures of $51.4 billion. The company bridged the gap with $41 billion raised from stock sales and new debt. Growth in the search advertising business, which accounts for about half of revenue, slowed to 12 percent, the lowest level in five quarters. AI answers in Google Search’s AI Mode are drawing usage away from the traditional advertising environment. Meta reported a negative $2.1 billion, despite advertising revenue rising 26 percent. Mark Zuckerberg described the new AI agent Muse as very compute-intensive to run.

Amazon posted a negative free cash flow of $42.4 billion over twelve months, compared to a positive $14.8 billion the previous year. The stock fell 7 percent following the earnings release, even as AWS grew by 39 percent. The cloud division’s order backlog rose to $578 billion, with new contracts including Meta, Nvidia, OpenAI, Salesforce, and Oracle. CEO Andy Jassy believes AWS can reach one trillion dollars in revenue in five to ten years and noted that Amazon installs an AI accelerator roughly every 14 seconds. According to a market researcher, spending on cloud services surged 55 percent to $168 billion in the third quarter. Providers like CoreWeave and Nebius, so-called Neocloud operators, are capturing market share from the established cloud giants.

Microsoft remained positive with $18.6 billion in free cash flow. Capital expenditures came in at $45.2 billion, below the announced $50 billion, which Microsoft attributed to delays among third-party computing capacity providers, whose shares subsequently declined. → 7 Wall St., Fortune, CIO Dive, Finimize, The Motley Fool, 7 Wall St., TechCrunch

Synthszr Take: The bill for $828 billion in capital expenditures ends up in every ETF savings plan. Index funds weight by market cap, meaning they carry the largest bet the most heavily. Alphabet lost around $500 billion on a single Thursday, and a noticeable portion of that is now missing from pension funds and retirement accounts. If this expansion is increasingly financed through new shares and debt, the risk trickles further down. Passive investing now means helping finance the AI buildout without ever having opted in.

OpenAI Stumbles on Capital Raise: Only $11 Billion Instead of $30 Billion

According to insiders, OpenAI has so far secured around $11 billion in commitments for its ongoing funding round. That is less than half of the originally targeted $30 billion. The company reportedly expects to raise additional capital in the coming weeks. According to the New York Post, the round is viewed as a precursor to a potential IPO. At the same time, CEO Sam Altman is facing criticism over comments regarding memory: on Stripe president John Collison’s YouTube show “Cheeky Pint,” he called it “quite terrible” that ordinary people tie up part of the world’s scarce DRAM in their personal devices. → New York Post

Synthszr Take: Branding a 5 percent utilization rate as waste invites investors to do the counter-calculation. In a round currently sitting at $11 billion of $30 billion, backers are asking precisely that: What is the utilization rate of the data centers being built with their capital? Altman’s lament about scarce DRAM also reveals that OpenAI is buying at scarcity prices. Ahead of an IPO, the CEO is putting his largest cost driver directly into the shop window.

DeepSeek Pulls It Off, Raising Nearly $16 Billion

DeepSeek closed a funding round of nearly $16 billion this month. Alongside major Chinese tech investors, several of the country’s most powerful state-backed institutions participated in the round, as reported by The Information citing details reviewed by the publication. According to the report, the investor roster reads like a who’s who of Chinese tech finance. → us.list-manage.com

Synthszr Take: Until now, DeepSeek essentially had one backer: its founder’s hedge-fund high-flyer. That independence was the lab’s quietest asset: research without quarterly pressures and without a political mandate. With nearly $16 billion, state-backed institutions are now sitting at the table, and capital of that scale buys influence. For companies outside China that rely on DeepSeek weights, this shifts the risk assessment. An open model backed by state shareholders will land on the exact same compliance pile as Huawei.

Anthropic Shows IPO Investors Slower Growth: Valuation Comes Under Pressure

According to The Information, Anthropic presented potential investors for its planned IPO with revenue growth that remains rapid but is slowing down. According to the report, the company shared the figures on Wednesday during a private investor meeting in San Francisco. The publicly accessible portion of the report does not disclose specific growth rates. → The Information

Synthszr Take: An IPO sells the slope of the curve, and Anthropic is currently showing investors that it is flattening. As long as growth was exploding, any valuation could be justified by the future. A flattening curve forces comparisons with mature software corporations, where sober multiples apply. The next two quarters thus become the actual roadshow. If they show stabilization, the valuation target holds. If they show further slowdown, the market will set the price—and OpenAI will instantly get a benchmark delivered alongside it.

Intel Plunges by a Third, Blackstone Warns of Junk Rating for Anthropic

The Nasdaq dropped 2.9 percent on Friday according to StrictlyVC, as investors increasingly doubt enormous AI capital expenditures. Intel plummeted 29 percent after missing expectations, marking its worst trading day since the 1970s. Other chip stocks also fell significantly: Qualcomm dropped about 11 percent, Nvidia fell 4 percent. Alphabet continued its losses after its shares had already fallen more than 12 percent on Thursday. That same week saw a warning from Blackstone: According to the Financial Times, the financial investor notified its investors that Anthropic could receive a high-yield or “junk” rating after its IPO. → StrictlyVC

Synthszr Take: A junk rating turns Anthropic’s data centers into an interest rate issue. Many pension funds and insurance companies are not allowed to hold high-yield bonds, meaning the very capital that otherwise cheaply finances long-term infrastructure disappears. What remains are more expensive creditors with shorter patience. Every percentage point in premium is then lost to compute, and compute is Anthropic’s growth engine. Alphabet pays for its data centers out of cash flow, even after dropping 12 percent in a single day. The gap relative to the hyperscalers is thus widening right on the balance sheet.

Sam Altman Calls PC RAM a “Quite Terrible” Waste

OpenAI CEO Sam Altman considers large amounts of RAM in personal computers to be an inefficient use of the globally scarce DRAM supply. In his view, cloud data centers could utilize this memory far more efficiently. The statement was made on the “Cheeky Pint” podcast during a conversation with Stripe co-founder John Collison. The prompt was whether locally run open-source models could noticeably reduce the need for data centers. Altman disagreed: He called it “quite terrible” for everyone to have a slice of the limited DRAM supply sitting in their own computer, which he estimated is used only about five percent of the time. According to Wccftech, he did not provide evidence for this figure. → Wccftech

Synthszr Take: Altman’s five-percent utilization is a number without evidence, but an argument with impact. Samsung has already sold out its HBM supply for 2027 and locked up 60 to 70 percent of capacity in long-term contracts. PC manufacturers and end consumers are competing for the remainder; in Japan, individual DDR5 kits cost up to 761 percent more than a year ago. Scarcity is pre-sorting the market: Local AI becomes expensive before it becomes widespread. The cloud wins the efficiency debate via the price list.

DeepSeek Open-Sources Almost Its Entire Agent Stack, Turing Post Tests What Holds Up

Alongside its model weights, DeepSeek has published an inference engine, an agent harness, and the underlying libraries—all the way down to inter-chip data transport. A guide by Alyona Vert and Ksenia Se in the Turing Post examines how much of this stack can be used to build custom agents. The starting point is DeepSeek-V4.1-Flash, released on September 10, 2026, whose updated weights have been available under the MIT license since October 17. According to DeepSeek, the model’s global KV cache occupies 890 bytes per token, roughly a quarter of the V4 Flash footprint. Architecturally, the model combines 552 billion backbone parameters with 196 billion parameters in a learned lookup store called Engram. → 🔳 Turing Post

Synthszr Take: DeepInfer powers DeepSeek’s own production; the public version can only compute offline. That gap is where the real work lies. An agent service thrives on scheduling, batching, load spikes, and a KV cache that must be managed under live traffic. The MIT license doesn’t bundle that part. Teams lacking in-house inference expertise inevitably return to DeepSeek’s API, leaving the open weights to serve merely as a bargaining chip. The stack is open, but operations remain in Hangzhou.

Anthropic Scraps Claude’s Five-Hour Limit, but Only for Ultra Subscribers

Anthropic is lifting the five-hour limit for Claude sessions for verified users on the Ultra tier. This was reported by The New Stack and relayed via the newsletter Techpresso. Previously, usage was capped in rolling five-hour windows. Users who exhausted their allowance had to wait until the next window opened. The change primarily impacts Claude Code, allowing sessions to run longer uninterrupted. → Techpresso

Synthszr Take: Anthropic is now putting a price tag on continuity. The weekly limit remains, meaning Ultra customers don’t get a single extra minute of compute time; they are merely allowed to consume their quota all at once. For agentic coding, that is the exact bottleneck, because an agent loses its workflow after a forced pause. Every tier below this now offers a tool with a built-in forced break every five hours. In the mid-range pricing tier, OpenAI and Cursor now find an open flank that Anthropic created itself.

Meta Tests AI Fashion App in Southeast Asia That Builds Outfits from Instagram Photos

According to TechCrunch, Meta has quietly launched the AI fashion app Style Sketch, which puts together outfit recommendations from an existing wardrobe. The app connects to Instagram and generates a virtual closet from posted photos. Alternatively, garments can be captured via photo, short video, or camera roll upload. Style Sketch uses this to suggest combinations tailored to personal style, which users can try on via a “realistic model” customized to body shape and skin tone, according to the app description. For trips, the app generates a daily outfit plan based on destination, duration, and planned activities, taking the local weather forecast into account. → us.list-manage.com

Synthszr Take: The most expensive moment for any wardrobe app is onboarding: taking photos of shirt after shirt before the first recommendation appears. Alta and Google Photos make their users go through precisely this tedious busywork. Meta, on the other hand, had this handed to them over years, because every posted Instagram outfit is already an inventoried piece of clothing. As a result, a new user costs Meta almost nothing. Accordingly, the test is running quietly in three Southeast Asian markets; a flop costs little more than the App Store listing.

Data Center Boom Threatens Republican Dominance in Ohio and Texas

According to The Hill, Republican dominance in Ohio and Texas is coming under pressure due to the construction boom in data centers. Both states are among the sought-after locations for new data centers being built to operate AI applications. According to the publication, the analysis focuses on growing public resentment over the consequences of these developments, including rising electricity prices. The Hill frames the development as a potential risk for the party ahead of the 2026 midterms. Ohio and Texas have long been considered reliable Republican strongholds.

Synthszr Take: The tax revenue from a data center appears once in a press release; the electricity bill arrives every month. This cadence is currently upending the locational logic in Ohio and Texas. When even Republican strongholds revolt against rising prices, public acceptance becomes a cost factor in hyperscalers' business cases. The next development will therefore need to bring its own power plants, higher industrial tariffs, or grid contributions. Compute is becoming more expensive—politically, before it does technically.

S&P Lowers Meta’s Credit Outlook, Doubting AI Monetization

On October 30, 2026, S&P Global Ratings lowered its outlook for Meta Platforms from stable to negative while affirming its AA- issuer rating. The move was justified by uncertainty over massive capital expenditures and negative free cash flow. The rating agency wrote that Meta’s ability to monetize its AI investments is “more limited than that of its competitors.” AA- is the fourth-highest tier on S&P’s scale. A negative outlook serves as a formal warning that a downgrade becomes possible if the agency’s projections materialize.

At the core of the decision is a significantly higher capex forecast. For 2026, S&P projects $140 billion, the midpoint of the $135 billion to $145 billion range Meta cited on October 21. For 2027, the agency now expects more than $225 billion, up from a previous $163 billion. For 2028, it projects $252 billion instead of $178 billion. S&P also cited the fact that Meta has not yet disclosed its spending plans for 2027. Capex growth is expected to flatten starting in 2028.

This results in free operating cash flow deficits of $22 billion for 2026, $71 billion for 2027, and $39 billion for 2028. Previously, S&P had expected a return to positive territory by 2028. The agency estimates that Meta will finance the 2028 deficit entirely with debt. The adjusted leverage ratio stood at around 0.1x in 2025. S&P expects 0.3x at the end of 2026, 1.1x in 2027, and 1.3x in 2028 as the new leases commence. The threshold for the current rating is 1.0x. The primary drivers are data center leases, which S&P treats like debt. Meta’s future lease obligations rose to $355.7 billion by the end of September, up from around $279 billion in the prior quarter. Long-term funded debt stood at $80.92 billion.

In the third quarter, Meta’s revenue increased by 26.1 percent to $64.6 billion, driven by a 25.9 percent increase in its advertising business. Capital expenditures reached $34.9 billion. Free cash flow came in at negative $2.12 billion, marking its first negative print in more than a decade. Added to that was a $9.89 billion charge stemming from a settlement with state attorneys general. Meta aims to recoup AI spending through improved advertising tools and agentic apps like Muse. → PPC Land, 7 Wall St., implicator, CNBC, The Information, Channel NewsAsia, Reuters, Finimize, Crypto Briefing

Synthszr Take: Meta has only one customer for its data centers: itself. Google, Microsoft, and Amazon rent compute to third parties; Meta must recoup every dollar through advertising. The 25.9 percent ad growth proves little on that front as long as no one breaks out the share driven by AI tools. That number is missing, as is the capex plan for 2027. Until Meta delivers both, S&P is bracing for the worst—and $355.7 billion in lease commitments gives the rating agency plenty of ammunition.

Anthropic’s Claude Duet Tested Over One Week Against Muse, Instinct, Dots, and Grok Bot

Anthropic has entered the proactive AI agent market with Claude Duet. Product analyst Aakash Gupta used the app for a week alongside Muse, Instinct, Dots, and Grok Bot. Anthropic unveiled Duet as a Research Preview on October 21, with the announcement amassing 11.2 million views. Gupta categorizes the app within a space that emerged this year: agents that reach out to users proactively rather than waiting for an input. Until now, it had been an open question how Anthropic would respond in this field.

According to Anthropic, Duet runs on Opus 5.5. The app delivers a briefing every morning, outlines next steps, and only chimes back in when a decision needs to be made. Tasks are assigned via text or voice, such as scheduling a meeting or building a dashboard. Deliverables are returned for review as a document, spreadsheet, or presentation. Memory, skills, and connected apps from the Claude account are available from the very first message within a single chat thread. As business use cases, the provider highlights tracking down outstanding invoices complete with draft dunning notices, preparing payroll, and finalizing contracts.

The agent continues working on a dedicated machine in the cloud while the user is away. Before taking sensitive actions like sending messages or making payments, it requests approval, according to the company. Approvals, logins, and security checks are designed to run within this cloud environment so sensitive data stays there. Users can see what the agent has completed, what it is currently doing, and intervene to steer it.

Access is initially available only via a waitlist for Pro, Max, and Ultra subscribers in the US, excluding Team and Enterprise plans. Duet shares usage limits with the rest of the Claude offering. The agent’s background checks also count toward this quota; larger tasks like a financial model require correspondingly more model calls. The app is available as a standalone client on iOS and Android, as well as in the browser.

The launch is accompanied by Anthropic’s prospectus. It reports a net loss of $80.7 billion for the first half of 2026, including $72.4 billion in non-cash charges from the revaluation of convertible notes, derivatives, and SAFE agreements. Operationally, the company reports a profit of $974 million for the second quarter, following an operating loss of $1.7 billion in the first quarter. More than 11,900 enterprise customers pay over $100,000 per year, compared to around 950 a year earlier. The last private round valued Anthropic at $965 billion in May, while $1.5 to $2 trillion is expected for the IPO. At $1.5 trillion, Gupta calculates this to be roughly 23 times annualized revenue. → Aakash Gupta, Aakash Gupta

Synthszr Take: Opus 5.5 under the hood will do little to decide Duet’s fate. Among coding copilots, whoever interrupted the workflow least won out. Proactive agents intensify this pattern because they demand attention without being asked. Duet’s most important design decision is therefore unassuming: a morning briefing, followed by radio silence until the next decision. Muse, Instinct, and Dots are thus competing on tact, which is won through product design. With Duet, every unnecessary background check also costs additional quota.