AI News Today June 17 2026: Top 10 AI Stories

AI News Today: Top 10 AI Stories - June 17, 2026

OpenAI's actual 2025 numbers are now public, and they are more complicated than the company's public narrative suggested. Audited financial documents reviewed by Ed Zitron's newsletter and independently verified by the Financial Times show $34 billion in total spending against $13 billion in revenue, producing a net loss attributable to the company of $38.5 billion. Separately, Microsoft confirmed it has been routing GitHub traffic through Amazon Web Services after AI coding agents drove platform availability to 88.4 percent in June. Anthropic leaders flew to Washington on Monday for emergency talks over Fable 5, and the two sides remain split on how serious the underlying security issue actually is. And Japan's dominant taxi app Go debuted on the Tokyo Stock Exchange with a 21 percent first-day gain, the country's largest IPO of 2026.

Zero overlap with our June 1 through June 16 posts. Here are the 10 stories that define today.

1. OpenAI's Audited 2025 Financials: $34B Spent, $13B Earned, $38.5B Net Loss

Audited financial documents from OpenAI's 2025 fiscal year, reviewed by Ed Zitron's Where's Your Ed At newsletter and independently verified by the Financial Times, show the company spent approximately $34 billion in 2025 while generating $13 billion in revenue. The net loss attributable to the company was $38.53 billion, roughly 7.5 times the $5.09 billion it lost in 2024 on $3.7 billion in revenue.

The expense breakdown: approximately $19 billion on research and development and nearly $6 billion on sales and marketing, with the remainder across infrastructure, operations, and general and administrative costs. OpenAI spent $5.02 billion on inference with Microsoft Azure in the first half of 2025 alone. For the full period from calendar 2024 through Q3 2025, inference spend on Azure totaled $12.43 billion, covering only serving costs, not training. The total compute bill is larger still.

The headline loss figure requires careful reading. The $38.53 billion net loss includes a $41.55 billion one-time charge tied to OpenAI's conversion from a nonprofit to a for-profit public benefit corporation in 2025, which triggered fair-value revaluation of convertible interests and warrant liability. A person familiar with the matter told the FT that most of the increase in losses came from this non-cash accounting charge rather than underlying operating performance. However, even stripping out the non-cash charge, the company's core operating losses are expanding substantially as each new model requires more compute to train and more compute to serve.

The revenue story is genuine: $13.07 billion in 2025, with monthly revenue reaching approximately $2 billion by year-end, compared to $1 billion per quarter at the end of 2024. But costs grew faster than revenue every quarter. The company had just over $50 billion in assets at year-end, with almost half in cash, supported by the $122 billion funding round closed earlier in 2026. This is the financial profile OpenAI's S-1 will need to explain to public investors: extraordinary revenue growth alongside losses that, net of the one-time charge, are still running at a pace that makes profitability dependent on assumptions about AI agent monetisation that have not yet been demonstrated at scale.

2. Microsoft Borrows AWS Capacity to Keep GitHub Online as AI Agents Overwhelm the Platform

Microsoft confirmed on June 16, 2026 that it is routing GitHub traffic through Amazon Web Services after a surge in AI coding agent activity pushed the platform past the reliability thresholds its enterprise customers expect. The disclosure was first reported by Business Insider, citing two people familiar with the plans. A Microsoft spokesperson confirmed that GitHub is using multiple cloud providers but declined to name Amazon directly, saying: 'The incredible spike in agentic development that began late last year has tested our infrastructure limits.'

The numbers behind the crisis: GitHub COO Kyle Daigle confirmed in April that the platform was processing 275 million commits per week, on pace for 14 billion in all of 2026, up from 1 billion in all of 2025. GitHub Actions weekly compute minutes grew from 500 million in 2023 to 2.1 billion in a single week in early 2026. AI agent-opened pull requests surged from 4 million in September 2025 to more than 17 million by March 2026. The platform logged nine service-degrading incidents in May 2026 alone and availability dropped to approximately 88.4 percent in June, well below the 99.9 percent enterprise SLA threshold.

Mitchell Hashimoto, co-founder of HashiCorp, captured the developer frustration on X: GitHub was 'no longer a place for serious work if it just blocks you out for hours per day, every day.' GitHub's own CTO Vlad Fedorov had acknowledged in February and March that the platform had breached its three-nines availability commitment, and disclosed in October 2025 that the company had begun executing a plan to increase capacity 10 times over.

The AWS arrangement is framed as a temporary operational measure while GitHub continues migrating its infrastructure to Azure by 2027. But the move carries a second-order implication. When Microsoft acquired GitHub for $7.5 billion in 2018, the strategic logic was that GitHub would become a natural on-ramp to Azure. Eight years later, GitHub's AI demand curve has exceeded Azure's ability to absorb it, and Microsoft's biggest cloud competitor is keeping its developer platform online. AI adoption is outpacing infrastructure planning even at the scale of the world's second-largest cloud provider.

3. Amazon CEO Andy Jassy's Warning Triggered the Fable 5 Government Shutdown

New reporting from the Wall Street Journal and Fortune on June 14-15, 2026 revealed a detail that had not been fully public: the immediate trigger for the US Department of Commerce's June 12 export control order on Fable 5 and Mythos 5 was a warning delivered to the White House by Amazon CEO Andy Jassy. Amazon, which is simultaneously Anthropic's largest financial backer and, through AWS, its primary infrastructure provider, had identified a jailbreak technique similar to the one Pliny the Liberator published publicly on June 10.

Jassy communicated Amazon's findings to members of the US administration, per the Wall Street Journal. This is the separate, private company claim that Axios had previously reported but not named. The sequence of events was: Amazon found a jailbreak and told the White House; the White House told the Commerce Department; the Commerce Department sent a letter to Anthropic on Friday, June 12, giving the company 90 minutes to restrict access. Anthropic CEO Dario Amodei then engaged in several calls with senior administration officials, per Politico, arguing the security bypass was narrow rather than a full jailbreak of the model's core safeguards.

The commercial relationship between Amazon and Anthropic is one of the most entangled in the AI industry. Amazon has invested approximately $8 billion in Anthropic across two tranches, making it the company's largest single investor. Anthropic committed to spending over $100 billion on AWS infrastructure in April 2026. Amazon deploys Claude through Amazon Bedrock as a key part of its enterprise AI offering. And yet it was Amazon's CEO who delivered the warning that led to the shutdown of Anthropic's most powerful models. The commercial partnership and the security concern were not in conflict from Amazon's perspective: protecting its government relationships and national security posture takes priority over the Anthropic investment relationship.

A source familiar with Anthropic told Fortune the company was given 90 minutes to implement the restrictions and received no previous communication of a specific national security threat. The government also had a suspicion, per Semafor, that a group with ties to China may have obtained access to the Mythos model, though it did not reveal how or which organisation was targeted. Reverse-engineering an AI model through distillation, where you use a powerful model's outputs to train a smaller derivative, is a recognised technique for extracting model capabilities without direct access to the weights. If Mythos-class intelligence was being distilled by a China-linked entity, the national security argument for restricting access becomes substantially more grounded than a narrow jailbreak debate.

4. Anthropic Leaders Fly to Washington for Fable 5 Talks - Both Sides Still Split on Risk

Anthropic executives flew to Washington DC on Monday, June 16, to meet with White House officials about the Fable 5 export control order. After the high-level talks, both sides remain split on the core question: how serious is the jailbreak risk? The White House AI and Crypto Czar David Sacks stated publicly on X that Anthropic 'refused to fix the issue,' and asked the rhetorical question of why, if Fable 5 was truly safe, the company had not already patched the vulnerability. Sacks also claimed that when Dario Amodei was informed of the jailbreak, he described it as not a serious risk.

Anthropic's position, as stated in its public statement and through company officials in the Washington meetings: the government approved Fable 5 before the global release. The vulnerability that has been identified is a narrow, non-universal technique that does not constitute a full jailbreak of the model's safety system. The company has argued consistently that applying the standard used to pull Fable 5, when the same information is accessible through other publicly deployed frontier models without any bypass, is logically inconsistent and would halt all new model deployments for every frontier AI provider if applied uniformly.

The standoff is being described by multiple observers as a test case for how aggressively Washington will regulate frontier AI models going forward. There is no indication yet of when, or whether, Fable 5 and Mythos 5 will return to public access. David Sacks's characterisation of Anthropic's position as refusing to fix the issue conflicts with Anthropic's characterisation of the vulnerability as non-patchable through conventional means, because the decomposition-and-recomposition technique the attacker used does not exploit a specific model flaw that a patch can address. It exploits the architecture of natural-language safety instructions, which is not fixable at the prompt engineering layer.

5. Anthropic Loses DC Circuit Stay but Court Orders Expedited Hearing on Pentagon Lawsuit

The US Court of Appeals for the DC Circuit denied Anthropic's request for an emergency stay of the Pentagon's supply chain risk designation in April 2026. The three-panel court found that Anthropic had not met the strict requirements for an emergency stay, but it granted the company's request for expedited treatment of the underlying case. Oral arguments were scheduled to begin May 19, 2026.

The Pentagon's designation of Anthropic as a supply chain risk in early March 2026, an action taken by Defense Secretary Pete Hegseth following the company's refusal to allow its AI to be used for mass domestic surveillance and fully autonomous lethal weapons, is the backdrop against which the June 12 export control order arrived. The two government actions are legally distinct but politically connected: Anthropic has been in open conflict with the national security establishment since March, and the Fable 5 shutdown arrived while that conflict was still in federal courts.

The California case produced a more favorable outcome for Anthropic. US District Judge Rita Lin issued a preliminary injunction in late March 2026 blocking the Pentagon designation, calling it 'classic' First Amendment retaliation and writing that the government's broad measures did not 'appear to be directed at the government's stated national security interests' but instead 'appear designed to punish Anthropic.' Acting Attorney General Todd Blanche described the DC Circuit stay denial as a 'resounding victory for military readiness.' The two federal courts reached opposite conclusions about whether the Pentagon's designation was lawful. The conflict is ongoing. The DC Circuit appeal was expedited, with oral arguments in May 2026, but no final ruling has been reported as of June 17.

6. Go Inc. Debuts on Tokyo Stock Exchange at Plus 21 Percent -- Japan's Largest IPO of 2026

Japan's most widely used taxi-hailing app, Go Inc., began trading on the Tokyo Stock Exchange Growth Market on June 16, 2026, surging 21 percent above its IPO price on debut day to close at approximately 2,910 yen per share. The company raised 88.6 billion yen, approximately $553 million, by pricing shares at 2,400 yen, the top of its indicated range of 2,350 to 2,400 yen. The offering was more than 25 times oversubscribed overall, with the international tranche alone more than 20 times covered by over 180 institutional investors including BlackRock and Wellington Management. Goldman Sachs and NTT Docomo backed the company.

Go is Japan's largest IPO of 2026, arriving at a thin time for the Tokyo market. There have been only 17 IPOs priced in Japan so far in 2026, the fewest since 2011, with total proceeds of just 144 billion yen, the lowest first-half figure since 2022. Go's clean investment story stood out in that environment: it is the dominant digital layer over Japan's existing taxi industry, built from the 2020 merger of JapanTaxi and DeNA's MOV business, and its growth path is straightforward. Revenue guidance for the fiscal year ending May 2027 is 40.8 billion yen, up approximately 30 percent from the prior year. Operating profit is guided to rise to 7 billion yen from 2.7 billion.

The AI angle: Go explicitly directed its IPO proceeds toward research and development for autonomous taxi services. The company is competing against Uber, Didi, and Sony-owned S.Ride for Japan's ride-hailing market, with the taxi-app category itself being a necessary infrastructure layer as autonomous vehicles enter commercial deployment. Japan's regulatory model for taxis, which kept Go as a partner to incumbents rather than a disruptor of them, gave it a more sustainable position than Uber's global playbook. For Japan's AI and autonomous vehicle ambitions, Go's successful listing provides the public capital to fund the technology roadmap.

7. Google Is Paying SpaceX $920 Million Per Month for Compute Because Its Own Cloud Cannot Keep Up

A previously disclosed deal, surfaced in infrastructure coverage this week, shows Google has agreed to pay SpaceX $920 million per month from October 2026 through June 2029 for access to compute capacity at xAI's Colossus data center. Google, which builds and operates one of the largest cloud computing networks in the world, described the arrangement as 'bridge capacity to meet surging customer demand' for its Gemini Enterprise AI platform that was 'even higher than we expected,' per TechCrunch's reporting.

The arrangement is extraordinary for what it reveals about the state of AI demand relative to infrastructure supply. Google has invested hundreds of billions of dollars in data centers and custom TPU chips. It designs and operates its own silicon through the Tensor Processing Unit line. It has an internal infrastructure team that is among the most technically sophisticated in the world. And it is still paying a competitor's data center $920 million per month -- more than $11 billion per year -- because it cannot build new capacity fast enough to meet what its own customers are already asking for.

This is the same market dynamic that drove Microsoft to use AWS for GitHub capacity. The AI demand curve has outrun the infrastructure planning cycle of even the largest technology companies simultaneously. For smaller organisations asking whether to commit to multi-year cloud AI contracts, the supply constraint is a meaningful factor: the capacity you are being offered today was built based on demand forecasts from 12 to 24 months ago. Demand has accelerated beyond those forecasts, and even the cloud providers are managing the gap through cross-competitor arrangements. Planning assumptions about AI infrastructure availability and pricing should account for continued constraint into 2027.

8. The AI Infrastructure Capacity Crisis: Even Cloud Giants Are Running Out of Runway

The week of June 16, 2026 has produced a cluster of stories that, taken together, describe a structural AI infrastructure capacity crisis that is broader than any individual company's problem. Microsoft is routing GitHub through AWS. Google is paying SpaceX $920 million per month for compute. Anthropic is paying SpaceX $1.25 billion per month for the Colossus 1 facility. NVIDIA's Trainium chips at Amazon are sold out through 2028. Oracle has a $638 billion backlog of committed AI infrastructure contracts it is racing to build out. The common thread: AI demand is consuming infrastructure faster than it can be built.

The constraint is multi-layered. Physical data center construction takes 18 to 36 months from site selection to operation. NVIDIA's Blackwell and upcoming Vera Rubin chip supply is constrained by TSMC's advanced node manufacturing capacity, which is itself the subject of export control negotiations between the US, Taiwan, and China. Power infrastructure for high-density AI data centers draws 40 to 100 kilowatts per rack, compared to 10 to 15 kilowatts for conventional cloud computing, requiring utility-scale power upgrades that take years and trigger the kind of rate case disputes documented in Arizona. Goldman Sachs estimates $7.6 trillion in cumulative AI capex from 2026 to 2031 will be required to meet projected demand -- equivalent to approximately one quarter of annual US GDP.

For developers and enterprise AI teams, the practical implications are two-fold. First, AI model pricing is unlikely to fall significantly over the next 12 to 18 months because the infrastructure cost structure is under rising pressure, not falling pressure. Second, AI API reliability is under more stress than the public incident reports from providers typically acknowledge. GitHub at 88.4 percent availability is the most visible example. But the same compute scarcity that is driving infrastructure partnerships across the industry also means that every model provider is managing capacity constraints that can produce elevated latency, request queuing, and service degradation under peak load.

9. OpenAI's IPO Story Gets Harder: $38.5B Net Loss Creates a Complex S-1 Disclosure

OpenAI's audited 2025 financials arrive at a particularly difficult moment: the company is preparing to file its public S-1 with the SEC for a September 2026 listing targeting a $1 trillion valuation. The $38.53 billion net loss disclosed in those audited documents, even adjusted for the one-time non-cash restructuring charge from the nonprofit-to-for-profit conversion, will be one of the most scrutinised line items in any S-1 filing in market history.

The most sophisticated argument in OpenAI's favour is also the most uncertain one: the company's revenue is growing exponentially, the model driving that revenue gets more capable with each generation, and enterprise AI agent products are just beginning to reach commercial deployment at scale. If AI agents become a standard part of enterprise operations across the sectors OpenAI is targeting, the revenue attached to each Codex installation or enterprise ChatGPT seat could compound significantly over the next three to five years.

The bear case, articulated by Ed Zitron in his analysis of the audited documents: 'The financial condition of OpenAI is deeply concerning. $38.53 billion in losses are astronomical, and far higher than most believed it would be. Losses also appear to be mounting year-over-year at a dramatic rate, and I'm not sure how this company finds a way toward any kind of sustainability or profitability.' Zitron noted that costs grew faster than revenue every quarter in 2025, and that the inference cost on Azure alone, $5.02 billion in the first half of the year, means the unit economics of serving ChatGPT at current scale are materially negative without an assumption that costs will fall or that significantly higher-margin products will replace the core consumer subscription.

Goldman Sachs and Morgan Stanley are leading the OpenAI offering. Their challenge: structuring an investor narrative around a company with extraordinary revenue growth and extraordinary losses in a public market environment where SPCX's first week has demonstrated that the appetite for AI-adjacent public equity is real, but the valuation discipline will be tested quickly by anyone who reads the actual numbers.

10. GitHub's Agent Economy: 17 Million AI Pull Requests Per Month and Rising

The GitHub infrastructure crisis provides a precise measurement of how fast AI agent adoption is compounding in software development. AI agent-opened pull requests on GitHub surged from 4 million in September 2025 to more than 17 million by March 2026, a more than fourfold increase in six months. That trajectory, if sustained, implies approaching 40 to 50 million AI agent pull requests per month by the end of 2026.

The distinction between a pull request opened by a human and one opened by an AI agent matters for infrastructure planning in a specific way. A human developer opens a pull request after hours or days of work, submitting it for review at a point in the workflow. An AI coding agent running on GitHub can open pull requests continuously, in parallel, across multiple repositories, around the clock. The compute, storage, and networking load per unit of output is similar to human pull requests, but the submission rate is not limited by human working hours, attention capacity, or fatigue. GitHub's infrastructure was sized for human developers. The agent economy has introduced a demand profile its capacity planning did not anticipate.

GitHub responded in April by moving all Copilot plans to usage-based billing, replacing premium request units with GitHub AI Credits calculated from token consumption. The billing change was a revenue mechanism, but it is also a demand management tool: metered pricing creates a feedback loop between usage volume and cost that flat-rate subscriptions do not. Whether it is enough to prevent the kind of sustained availability degradation the platform has experienced throughout the first half of 2026 remains to be seen. The AWS capacity addition is a temporary bridge, not a structural solution. The structural solution is Azure migration and internal capacity expansion, which GitHub has been executing since October 2025 on a timeline that has been consistently overtaken by demand growth.

Frequently Asked Questions

Q: How much did OpenAI spend in 2025?

Audited financial documents reviewed by Ed Zitron's Where's Your Ed At and independently verified by the Financial Times show OpenAI spent approximately $34 billion in 2025, including roughly $19 billion on research and development and nearly $6 billion on sales and marketing. The company generated $13.07 billion in revenue in 2025. The net loss attributable to OpenAI was $38.53 billion, largely inflated by a $41.55 billion one-time non-cash charge related to OpenAI's conversion from a nonprofit to a for-profit public benefit corporation. Source: Ed Zitron, Where's Your Ed At; Financial Times (June 15, 2026).

Q: Why is Microsoft using AWS for GitHub?

Microsoft confirmed on June 16, 2026 that it is routing GitHub traffic through Amazon Web Services after AI coding agents overwhelmed the platform's infrastructure. GitHub COO Kyle Daigle confirmed in April that the platform was processing 275 million commits per week, on pace for 14 billion in 2026 versus 1 billion in 2025. AI agent-opened pull requests grew from 4 million in September 2025 to 17 million by March 2026. GitHub logged nine service incidents in May and availability dropped to roughly 88.4 percent in June, well below the 99.9 percent enterprise SLA. Microsoft described the AWS arrangement as a temporary measure while GitHub continues migrating to Azure by 2027. Sources: Business Insider (June 16, 2026); AI Weekly (June 16, 2026); TechTimes (June 16, 2026).

Q: Who triggered the Fable 5 government shutdown?

Amazon CEO Andy Jassy communicated a jailbreak finding to White House officials, which triggered the June 12, 2026 US Commerce Department export control order pulling Fable 5 and Mythos 5 offline. Amazon had identified a vulnerability similar to the one published by Pliny the Liberator on June 10. The Commerce Department sent a letter to Anthropic on June 12 giving the company 90 minutes to restrict access. Anthropic CEO Dario Amodei then engaged in calls with senior administration officials arguing the bypass was narrow and not a full jailbreak. Sources: Wall Street Journal (June 14, 2026); Fortune (June 14, 2026); MLQ.ai; TechPolicy.Press.

Q: What happened at the Anthropic White House meeting on June 16?

Anthropic leaders flew to Washington DC on June 16 for high-level talks with White House officials over the Fable 5 export control order. After the meeting, both sides remained split on the core question of how serious the jailbreak risk is. White House AI and Crypto Czar David Sacks stated publicly that Anthropic refused to fix the issue and questioned why, if Fable 5 was safe, the vulnerability had not been patched. Anthropic maintains the vulnerability is narrow, non-universal, and that the government had approved Fable 5 before its global release. No restoration timeline for Fable 5 or Mythos 5 has been announced. Sources: BusinessToday (June 16, 2026); TechJuice (June 16, 2026); TechPolicy.Press.

Q: What is the Go Inc. IPO?

Go Inc. is Japan's most widely used taxi-hailing app, built from the 2020 merger of JapanTaxi and DeNA's MOV business. It raised 88.6 billion yen ($553 million) by pricing shares at 2,400 yen at the top of its range, valuing the company at 186 billion yen. Shares debuted on the Tokyo Stock Exchange Growth Market on June 16, 2026, surging 21 percent to 2,910 yen. The IPO was more than 25 times oversubscribed. Backed by Goldman Sachs and NTT Docomo, Go is Japan's largest IPO of 2026. Sources: Japan Times (June 16, 2026); Bloomberg (June 15, 2026); The Next Web (June 16, 2026).

Q: Why is Google paying SpaceX $920 million per month for compute?

Google agreed to pay SpaceX $920 million per month from October 2026 through June 2029 for access to xAI's Colossus data center compute capacity. Google described the deal as bridge capacity to meet Gemini Enterprise customer demand 'even higher than we expected.' The deal reveals that even the world's largest cloud operators face a gap between AI demand and their ability to build new infrastructure fast enough. The arrangement runs in parallel with Anthropic paying SpaceX $1.25 billion per month for Colossus 1 capacity and Microsoft using AWS for GitHub. Source: TechCrunch; Runtime Wire (June 16, 2026).

Q: What is the AI infrastructure capacity crisis?

Multiple major AI and cloud companies are simultaneously unable to meet AI compute demand with their own infrastructure, requiring cross-competitor capacity arrangements. Microsoft uses AWS for GitHub. Google pays SpaceX $920 million per month for compute. Anthropic pays SpaceX $1.25 billion per month for Colossus 1. NVIDIA Trainium chips at AWS are sold out through 2028. Oracle has $638 billion in committed AI infrastructure it is racing to build. Goldman Sachs projects $7.6 trillion in cumulative AI capex will be required from 2026 to 2031. Data center construction takes 18 to 36 months. Power upgrades for AI-density racks take years. The demand curve is outrunning the construction cycle.